diff --git a/README.md b/README.md
index 62ff566..45f8f98 100644
--- a/README.md
+++ b/README.md
@@ -16,6 +16,17 @@ Using UV, installing the packages and virtual environment is as simple as typing
 uv sync
 ```
 
+## Local LLM
+
+To run a LLM locally download https://lmstudio.ai
+When installing select developer mode, download a model (it will already suggest one) and run it (see developer window, status: running)
+
+copy the url at the top right and replace local_llm_url with it + v1/chat/completions.
+This + part might differ based on what model you choose.
+
+copy the model name in the module loaded and replace local_llm_modelL. In settings.
+
+
 ## Running
 To run the project (development server), execute the following command (while inside the root repository):
 
@@ -24,10 +35,16 @@ uv run fastapi dev src/control_backend/main.py
 ```
 
 ## Testing
-Testing happens automatically when opening a merge request to any branch. If you want to manually run the test suite, you can do so by running the following:
+Testing happens automatically when opening a merge request to any branch. If you want to manually run the test suite, you can do so by running the following for unit tests:
 
 ```bash
-uv run --only-group test pytest
+uv run --only-group test pytest test/unit
+```
+
+Or for integration tests:
+
+```bash
+uv run --group integration-test pytest test/integration
 ```
 
 ## GitHooks
diff --git a/pyproject.toml b/pyproject.toml
index 8299d0f..ee3ca08 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -7,6 +7,7 @@ requires-python = ">=3.13"
 dependencies = [
     "fastapi[all]>=0.115.6",
     "mlx-whisper>=0.4.3 ; sys_platform == 'darwin'",
+    "numpy>=2.3.3",
     "openai-whisper>=20250625",
     "pyaudio>=0.2.14",
     "pydantic>=2.12.0",
@@ -33,6 +34,7 @@ integration-test = [
     "soundfile>=0.13.1",
 ]
 test = [
+    "numpy>=2.3.3",
     "pytest>=8.4.2",
     "pytest-asyncio>=1.2.0",
     "pytest-cov>=7.0.0",
diff --git a/src/control_backend/agents/bdi/bdi_core.py b/src/control_backend/agents/bdi/bdi_core.py
index 1696303..6e5cdc0 100644
--- a/src/control_backend/agents/bdi/bdi_core.py
+++ b/src/control_backend/agents/bdi/bdi_core.py
@@ -1,9 +1,15 @@
 import logging
 
 import agentspeak
+from spade.behaviour import OneShotBehaviour
+from spade.message import Message
 from spade_bdi.bdi import BDIAgent
 
-from control_backend.agents.bdi.behaviours.belief_setter import BeliefSetter
+from control_backend.agents.bdi.behaviours.belief_setter import BeliefSetterBehaviour
+from control_backend.agents.bdi.behaviours.receive_llm_resp_behaviour import (
+    ReceiveLLMResponseBehaviour,
+)
+from control_backend.core.config import settings
 
 
 class BDICoreAgent(BDIAgent):
@@ -11,25 +17,52 @@ class BDICoreAgent(BDIAgent):
     This is the Brain agent that does the belief inference with AgentSpeak.
     This is a continous process that happens automatically in the background.
     This class contains all the actions that can be called from AgentSpeak plans.
-    It has the BeliefSetter behaviour.
+    It has the BeliefSetter behaviour and can aks and recieve requests from the LLM agent.
     """
 
-    logger = logging.getLogger("BDI Core")
+    logger = logging.getLogger("bdi_core_agent")
 
-    async def setup(self):
-        belief_setter = BeliefSetter()
-        self.add_behaviour(belief_setter)
+    async def setup(self) -> None:
+        """
+        Initializes belief behaviors and message routing.
+        """
+        self.logger.info("BDICoreAgent setup started")
+
+        self.add_behaviour(BeliefSetterBehaviour())
+        self.add_behaviour(ReceiveLLMResponseBehaviour())
+
+        self.logger.info("BDICoreAgent setup complete")
+
+    def add_custom_actions(self, actions) -> None:
+        """
+        Registers custom AgentSpeak actions callable from plans.
+        """
 
-    def add_custom_actions(self, actions):
         @actions.add(".reply", 1)
-        def _reply(agent, term, intention):
-            message = agentspeak.grounded(term.args[0], intention.scope)
-            self.logger.info(f"Replying to message: {message}")
-            reply = self._send_to_llm(message)
-            self.logger.info(f"Received reply: {reply}")
+        def _reply(agent: "BDICoreAgent", term, intention):
+            """
+            Sends text to the LLM (AgentSpeak action).
+            Example: .reply("Hello LLM!")
+            """
+            message_text = agentspeak.grounded(term.args[0], intention.scope)
+            self.logger.info("Reply action sending: %s", message_text)
 
+            self._send_to_llm(str(message_text))
             yield
 
-    def _send_to_llm(self, message) -> str:
-        """TODO: implement"""
-        return f"This is a reply to {message}"
+    def _send_to_llm(self, text: str):
+        """
+        Sends a text query to the LLM Agent asynchronously.
+        """
+
+        class SendBehaviour(OneShotBehaviour):
+            async def run(self) -> None:
+                msg = Message(
+                    to=settings.agent_settings.llm_agent_name + "@" + settings.agent_settings.host,
+                    body=text,
+                )
+
+                await self.send(msg)
+                self.agent.logger.info("Message sent to LLM: %s", text)
+
+        self.add_behaviour(SendBehaviour())
diff --git a/src/control_backend/agents/bdi/behaviours/belief_setter.py b/src/control_backend/agents/bdi/behaviours/belief_setter.py
index d36fe5e..2f64036 100644
--- a/src/control_backend/agents/bdi/behaviours/belief_setter.py
+++ b/src/control_backend/agents/bdi/behaviours/belief_setter.py
@@ -1,4 +1,3 @@
-import asyncio
 import json
 import logging
 
@@ -9,11 +8,10 @@ from spade_bdi.bdi import BDIAgent
 from control_backend.core.config import settings
 
 
-class BeliefSetter(CyclicBehaviour):
+class BeliefSetterBehaviour(CyclicBehaviour):
     """
     This is the behaviour that the BDI agent runs. This behaviour waits for incoming
-    message and processes it based on sender. Currently, it only waits for messages
-    containing beliefs from BeliefCollector and adds these to its KB.
+    message and processes it based on sender.
     """
 
     agent: BDIAgent
@@ -24,7 +22,6 @@ class BeliefSetter(CyclicBehaviour):
         if msg:
             self.logger.info(f"Received message {msg.body}")
             self._process_message(msg)
-        await asyncio.sleep(1)
 
     def _process_message(self, message: Message):
         sender = message.sender.node  # removes host from jid and converts to str
@@ -35,6 +32,7 @@ class BeliefSetter(CyclicBehaviour):
                 self.logger.debug("Processing message from belief collector.")
                 self._process_belief_message(message)
             case _:
+                self.logger.debug("Not the belief agent, discarding message")
                 pass
 
     def _process_belief_message(self, message: Message):
@@ -44,19 +42,25 @@ class BeliefSetter(CyclicBehaviour):
         match message.thread:
             case "beliefs":
                 try:
-                    beliefs: dict[str, list[list[str]]] = json.loads(message.body)
+                    beliefs: dict[str, list[str]] = json.loads(message.body)
                     self._set_beliefs(beliefs)
                 except json.JSONDecodeError as e:
                     self.logger.error("Could not decode beliefs into JSON format: %s", e)
             case _:
                 pass
 
-    def _set_beliefs(self, beliefs: dict[str, list[list[str]]]):
+    def _set_beliefs(self, beliefs: dict[str, list[str]]):
+        """Remove previous values for beliefs and update them with the provided values."""
         if self.agent.bdi is None:
             self.logger.warning("Cannot set beliefs, since agent's BDI is not yet initialized.")
             return
 
-        for belief, arguments_list in beliefs.items():
-            for arguments in arguments_list:
-                self.agent.bdi.set_belief(belief, *arguments)
-                self.logger.info("Set belief %s with arguments %s", belief, arguments)
+        # Set new beliefs (outdated beliefs are automatically removed)
+        for belief, arguments in beliefs.items():
+            self.agent.bdi.set_belief(belief, *arguments)
+
+            # Special case: if there's a new user message, flag that we haven't responded yet
+            if belief == "user_said":
+                self.agent.bdi.set_belief("new_message")
+
+            self.logger.info("Set belief %s with arguments %s", belief, arguments)
diff --git a/src/control_backend/agents/bdi/behaviours/receive_llm_resp_behaviour.py b/src/control_backend/agents/bdi/behaviours/receive_llm_resp_behaviour.py
new file mode 100644
index 0000000..dc6e862
--- /dev/null
+++ b/src/control_backend/agents/bdi/behaviours/receive_llm_resp_behaviour.py
@@ -0,0 +1,28 @@
+import logging
+
+from spade.behaviour import CyclicBehaviour
+
+from control_backend.core.config import settings
+
+
+class ReceiveLLMResponseBehaviour(CyclicBehaviour):
+    """
+    Adds behavior to receive responses from the LLM Agent.
+    """
+
+    logger = logging.getLogger("BDI/LLM Reciever")
+
+    async def run(self):
+        msg = await self.receive(timeout=2)
+        if not msg:
+            return
+
+        sender = msg.sender.node
+        match sender:
+            case settings.agent_settings.llm_agent_name:
+                content = msg.body
+                self.logger.info("Received LLM response: %s", content)
+                # Here the BDI can pass the message back as a response
+            case _:
+                self.logger.debug("Not from the llm, discarding message")
+                pass
diff --git a/src/control_backend/agents/bdi/behaviours/text_belief_extractor.py b/src/control_backend/agents/bdi/behaviours/text_belief_extractor.py
new file mode 100644
index 0000000..913bc44
--- /dev/null
+++ b/src/control_backend/agents/bdi/behaviours/text_belief_extractor.py
@@ -0,0 +1,100 @@
+import asyncio
+import json
+import logging
+
+from spade.behaviour import CyclicBehaviour
+from spade.message import Message
+
+from control_backend.core.config import settings
+
+
+class BeliefFromText(CyclicBehaviour):
+    logger = logging.getLogger("Belief From Text")
+
+    # TODO: LLM prompt nog hardcoded
+    llm_instruction_prompt = """
+    You are an information extraction assistent for a BDI agent. 
+    Your task is to extract values from a user's text to bind a list of ungrounded beliefs. Rules:
+    You will receive a JSON object with "beliefs" 
+    (a list of ungrounded AgentSpeak beliefs) and "text" (user's transcript).
+    Analyze the text to find values that sematically match the variables (X,Y,Z) in the beliefs.
+    A single piece of text might contain multiple instances that match a belief.
+    Respond ONLY with a single JSON object.
+    The JSON object's keys should be the belief functors (e.g., "weather").
+    The value for each key must be a list of lists.
+    Each inner list must contain the extracted arguments
+    (as strings) for one instance of that belief.
+    CRITICAL: If no information in the text matches a belief,
+    DO NOT include that key in your response.
+    """
+
+    # on_start agent receives message containing the beliefs to look out
+    #  for and sets up the LLM with instruction prompt
+    # async def on_start(self):
+    #    msg = await self.receive(timeout=0.1)
+    #    self.beliefs = dict uit message
+    #    send instruction prompt to LLM
+
+    beliefs: dict[str, list[str]]
+    beliefs = {"mood": ["X"], "car": ["Y"]}
+
+    async def run(self):
+        msg = await self.receive(timeout=0.1)
+        if msg:
+            sender = msg.sender.node
+            match sender:
+                case settings.agent_settings.transcription_agent_name:
+                    self.logger.info("Received text from transcriber.")
+                    await self._process_transcription_demo(msg.body)
+                case _:
+                    self.logger.info("Received message from other agent.")
+                    pass
+        await asyncio.sleep(1)
+
+    async def _process_transcription(self, text: str):
+        text_prompt = f"Text: {text}"
+
+        beliefs_prompt = "These are the beliefs to be bound:\n"
+        for belief, values in self.beliefs.items():
+            beliefs_prompt += f"{belief}({', '.join(values)})\n"
+
+        prompt = text_prompt + beliefs_prompt
+        self.logger.info(prompt)
+        # prompt_msg = Message(to="LLMAgent@whatever")
+        # response = self.send(prompt_msg)
+
+        # Mock response; response is beliefs in JSON format, it parses do dict[str,list[list[str]]]
+        response = '{"mood": [["happy"]]}'
+        # Verify by trying to parse
+        try:
+            json.loads(response)
+            belief_message = Message(
+                to=settings.agent_settings.bdi_core_agent_name + "@" + settings.agent_settings.host,
+                body=response,
+            )
+            belief_message.thread = "beliefs"
+
+            await self.send(belief_message)
+            self.logger.info("Sent beliefs to BDI.")
+        except json.JSONDecodeError:
+            # Parsing failed, so the response is in the wrong format, log warning
+            self.logger.warning("Received LLM response in incorrect format.")
+
+    async def _process_transcription_demo(self, txt: str):
+        """
+        Demo version to process the transcription input to beliefs. For the demo only the belief
+        'user_said' is relevant, so this function simply makes a dict with key: "user_said",
+        value: txt and passes this to the Belief Collector agent.
+        """
+        belief = {"beliefs": {"user_said": [txt]}, "type": "belief_extraction_text"}
+        payload = json.dumps(belief)
+        belief_msg = Message(
+            to=settings.agent_settings.belief_collector_agent_name
+            + "@"
+            + settings.agent_settings.host,
+            body=payload,
+        )
+        belief_msg.thread = "beliefs"
+
+        await self.send(belief_msg)
+        self.logger.info("Sent beliefs to Belief Collector.")
diff --git a/src/control_backend/agents/bdi/rules.asl b/src/control_backend/agents/bdi/rules.asl
index 41660a4..0001d3c 100644
--- a/src/control_backend/agents/bdi/rules.asl
+++ b/src/control_backend/agents/bdi/rules.asl
@@ -1,3 +1,3 @@
-+user_said(Message) : not responded <-
-  +responded;
++new_message : user_said(Message) <-
+  -new_message;
   .reply(Message).
diff --git a/src/control_backend/agents/bdi/text_extractor.py b/src/control_backend/agents/bdi/text_extractor.py
new file mode 100644
index 0000000..ff9ad58
--- /dev/null
+++ b/src/control_backend/agents/bdi/text_extractor.py
@@ -0,0 +1,9 @@
+from spade.agent import Agent
+
+from control_backend.agents.bdi.behaviours.text_belief_extractor import BeliefFromText
+
+
+class TBeliefExtractor(Agent):
+    async def setup(self):
+        self.b = BeliefFromText()
+        self.add_behaviour(self.b)
diff --git a/test/__init__.py b/src/control_backend/agents/belief_collector/__init__.py
similarity index 100%
rename from test/__init__.py
rename to src/control_backend/agents/belief_collector/__init__.py
diff --git a/src/control_backend/agents/belief_collector/behaviours/continuous_collect.py b/src/control_backend/agents/belief_collector/behaviours/continuous_collect.py
new file mode 100644
index 0000000..ada1c7a
--- /dev/null
+++ b/src/control_backend/agents/belief_collector/behaviours/continuous_collect.py
@@ -0,0 +1,117 @@
+import json
+import logging
+
+from spade.agent import Message
+from spade.behaviour import CyclicBehaviour
+
+from control_backend.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+class ContinuousBeliefCollector(CyclicBehaviour):
+    """
+    Continuously collects beliefs/emotions from extractor agents:
+    Then we send a unified belief packet to the BDI agent.
+    """
+
+    async def run(self):
+        msg = await self.receive(timeout=0.1)  # Wait for 0.1s
+        if msg:
+            await self._process_message(msg)
+
+    async def _process_message(self, msg: Message):
+        sender_node = self._sender_node(msg)
+
+        # Parse JSON payload
+        try:
+            payload = json.loads(msg.body)
+        except Exception as e:
+            logger.warning(
+                "BeliefCollector: failed to parse JSON from %s. Body=%r Error=%s",
+                sender_node,
+                msg.body,
+                e,
+            )
+            return
+
+        msg_type = payload.get("type")
+
+        # Prefer explicit 'type' field
+        if msg_type == "belief_extraction_text" or sender_node == "belief_text_agent_mock":
+            logger.info(
+                "BeliefCollector: message routed to _handle_belief_text (sender=%s)", sender_node
+            )
+            await self._handle_belief_text(payload, sender_node)
+        # This is not implemented yet, but we keep the structure for future use
+        elif msg_type == "emotion_extraction_text" or sender_node == "emo_text_agent_mock":
+            logger.info(
+                "BeliefCollector: message routed to _handle_emo_text (sender=%s)", sender_node
+            )
+            await self._handle_emo_text(payload, sender_node)
+        else:
+            logger.info(
+                "BeliefCollector: unrecognized message (sender=%s, type=%r). Ignoring.",
+                sender_node,
+                msg_type,
+            )
+
+    @staticmethod
+    def _sender_node(msg: Message) -> str:
+        """
+        Extracts the 'node' (localpart) of the sender JID.
+        E.g., 'agent@host/resource' -> 'agent'
+        """
+        s = str(msg.sender) if msg.sender is not None else "no_sender"
+        return s.split("@", 1)[0] if "@" in s else s
+
+    async def _handle_belief_text(self, payload: dict, origin: str):
+        """
+        Expected payload:
+         {
+            "type": "belief_extraction_text",
+            "beliefs": {"user_said": ["hello"","Can you help me?",
+            "stop talking to me","No","Pepper do a dance"]}
+
+        }
+
+        """
+        beliefs = payload.get("beliefs", {})
+
+        if not beliefs:
+            logger.info("BeliefCollector: no beliefs to process.")
+            return
+
+        if not isinstance(beliefs, dict):
+            logger.warning("BeliefCollector: 'beliefs' is not a dict: %r", beliefs)
+            return
+
+        if not all(isinstance(v, list) for v in beliefs.values()):
+            logger.warning("BeliefCollector: 'beliefs' values are not all lists: %r", beliefs)
+            return
+
+        logger.info("BeliefCollector: forwarding %d beliefs.", len(beliefs))
+        for belief_name, belief_list in beliefs.items():
+            for belief in belief_list:
+                logger.info("  - %s %s", belief_name, str(belief))
+
+        await self._send_beliefs_to_bdi(beliefs, origin=origin)
+
+    async def _handle_emo_text(self, payload: dict, origin: str):
+        """TODO: implement (after we have emotional recogntion)"""
+        pass
+
+    async def _send_beliefs_to_bdi(self, beliefs: list[str], origin: str | None = None):
+        """
+        Sends a unified belief packet to the BDI agent.
+        """
+        if not beliefs:
+            return
+
+        to_jid = f"{settings.agent_settings.bdi_core_agent_name}@{settings.agent_settings.host}"
+
+        msg = Message(to=to_jid, sender=self.agent.jid, thread="beliefs")
+        msg.body = json.dumps(beliefs)
+
+        await self.send(msg)
+        logger.info("BeliefCollector: sent %d belief(s) to BDI at %s", len(beliefs), to_jid)
diff --git a/src/control_backend/agents/belief_collector/belief_collector.py b/src/control_backend/agents/belief_collector/belief_collector.py
new file mode 100644
index 0000000..8558242
--- /dev/null
+++ b/src/control_backend/agents/belief_collector/belief_collector.py
@@ -0,0 +1,15 @@
+import logging
+
+from spade.agent import Agent
+
+from .behaviours.continuous_collect import ContinuousBeliefCollector
+
+logger = logging.getLogger(__name__)
+
+
+class BeliefCollectorAgent(Agent):
+    async def setup(self):
+        logger.info("BeliefCollectorAgent starting (%s)", self.jid)
+        # Attach the continuous collector behaviour (listens and forwards to BDI)
+        self.add_behaviour(ContinuousBeliefCollector())
+        logger.info("BeliefCollectorAgent ready.")
diff --git a/src/control_backend/agents/llm/llm.py b/src/control_backend/agents/llm/llm.py
new file mode 100644
index 0000000..c3c17ab
--- /dev/null
+++ b/src/control_backend/agents/llm/llm.py
@@ -0,0 +1,123 @@
+"""
+LLM Agent module for routing text queries from the BDI Core Agent to a local LLM
+service and returning its responses back to the BDI Core Agent.
+"""
+
+import logging
+from typing import Any
+
+import httpx
+from spade.agent import Agent
+from spade.behaviour import CyclicBehaviour
+from spade.message import Message
+
+from control_backend.agents.llm.llm_instructions import LLMInstructions
+from control_backend.core.config import settings
+
+
+class LLMAgent(Agent):
+    """
+    Agent responsible for processing user text input and querying a locally
+    hosted LLM for text generation. Receives messages from the BDI Core Agent
+    and responds with processed LLM output.
+    """
+
+    logger = logging.getLogger("llm_agent")
+
+    class ReceiveMessageBehaviour(CyclicBehaviour):
+        """
+        Cyclic behaviour to continuously listen for incoming messages from
+        the BDI Core Agent and handle them.
+        """
+
+        async def run(self):
+            """
+            Receives SPADE messages and processes only those originating from the
+            configured BDI agent.
+            """
+            msg = await self.receive(timeout=1)
+            if not msg:
+                return
+
+            sender = msg.sender.node
+            self.agent.logger.info(
+                "Received message: %s from %s",
+                msg.body,
+                sender,
+            )
+
+            if sender == settings.agent_settings.bdi_core_agent_name:
+                self.agent.logger.debug("Processing message from BDI Core Agent")
+                await self._process_bdi_message(msg)
+            else:
+                self.agent.logger.debug("Message ignored (not from BDI Core Agent)")
+
+        async def _process_bdi_message(self, message: Message):
+            """
+            Forwards user text to the LLM and replies with the generated text.
+            """
+            user_text = message.body
+            llm_response = await self._query_llm(user_text)
+            await self._reply(llm_response)
+
+        async def _reply(self, msg: str):
+            """
+            Sends a response message back to the BDI Core Agent.
+            """
+            reply = Message(
+                to=settings.agent_settings.bdi_core_agent_name + "@" + settings.agent_settings.host,
+                body=msg,
+            )
+            await self.send(reply)
+            self.agent.logger.info("Reply sent to BDI Core Agent")
+
+        async def _query_llm(self, prompt: str) -> str:
+            """
+            Sends a chat completion request to the local LLM service.
+
+            :param prompt: Input text prompt to pass to the LLM.
+            :return: LLM-generated content or fallback message.
+            """
+            async with httpx.AsyncClient(timeout=120.0) as client:
+                # Example dynamic content for future (optional)
+
+                instructions = LLMInstructions()
+                developer_instruction = instructions.build_developer_instruction()
+
+                response = await client.post(
+                    settings.llm_settings.local_llm_url,
+                    headers={"Content-Type": "application/json"},
+                    json={
+                        "model": settings.llm_settings.local_llm_model,
+                        "messages": [
+                            {"role": "developer", "content": developer_instruction},
+                            {"role": "user", "content": prompt},
+                        ],
+                        "temperature": 0.3,
+                    },
+                )
+
+                try:
+                    response.raise_for_status()
+                    data: dict[str, Any] = response.json()
+                    return (
+                        data.get("choices", [{}])[0]
+                        .get("message", {})
+                        .get("content", "No response")
+                    )
+                except httpx.HTTPError as err:
+                    self.agent.logger.error("HTTP error: %s", err)
+                    return "LLM service unavailable."
+                except Exception as err:
+                    self.agent.logger.error("Unexpected error: %s", err)
+                    return "Error processing the request."
+
+    async def setup(self):
+        """
+        Sets up the SPADE behaviour to filter and process messages from the
+        BDI Core Agent.
+        """
+        self.logger.info("LLMAgent setup complete")
+
+        behaviour = self.ReceiveMessageBehaviour()
+        self.add_behaviour(behaviour)
diff --git a/src/control_backend/agents/llm/llm_instructions.py b/src/control_backend/agents/llm/llm_instructions.py
new file mode 100644
index 0000000..9636d88
--- /dev/null
+++ b/src/control_backend/agents/llm/llm_instructions.py
@@ -0,0 +1,44 @@
+class LLMInstructions:
+    """
+    Defines structured instructions that are sent along with each request
+    to the LLM to guide its behavior (norms, goals, etc.).
+    """
+
+    @staticmethod
+    def default_norms() -> str:
+        return """
+        Be friendly and respectful.
+        Make the conversation feel natural and engaging.
+        """.strip()
+
+    @staticmethod
+    def default_goals() -> str:
+        return """
+        Try to learn the user's name during conversation.
+        """.strip()
+
+    def __init__(self, norms: str | None = None, goals: str | None = None):
+        self.norms = norms if norms is not None else self.default_norms()
+        self.goals = goals if goals is not None else self.default_goals()
+
+    def build_developer_instruction(self) -> str:
+        """
+        Builds a multi-line formatted instruction string for the LLM.
+        Includes only non-empty structured fields.
+        """
+        sections = [
+            "You are a Pepper robot engaging in natural human conversation.",
+            "Keep responses between 1–5 sentences, unless instructed otherwise.\n",
+        ]
+
+        if self.norms:
+            sections.append("Norms to follow:")
+            sections.append(self.norms)
+            sections.append("")
+
+        if self.goals:
+            sections.append("Goals to reach:")
+            sections.append(self.goals)
+            sections.append("")
+
+        return "\n".join(sections).strip()
diff --git a/src/control_backend/agents/mock_agents/__init__.py b/src/control_backend/agents/mock_agents/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/src/control_backend/agents/mock_agents/belief_text_mock.py b/src/control_backend/agents/mock_agents/belief_text_mock.py
new file mode 100644
index 0000000..769b263
--- /dev/null
+++ b/src/control_backend/agents/mock_agents/belief_text_mock.py
@@ -0,0 +1,43 @@
+import json
+
+from spade.agent import Agent
+from spade.behaviour import OneShotBehaviour
+from spade.message import Message
+
+from control_backend.core.config import settings
+
+
+class BeliefTextAgent(Agent):
+    class SendOnceBehaviourBlfText(OneShotBehaviour):
+        async def run(self):
+            to_jid = (
+                f"{settings.agent_settings.belief_collector_agent_name}"
+                f"@{settings.agent_settings.host}"
+            )
+
+            # Send multiple beliefs in one JSON payload
+            payload = {
+                "type": "belief_extraction_text",
+                "beliefs": {
+                    "user_said": [
+                        "hello test",
+                        "Can you help me?",
+                        "stop talking to me",
+                        "No",
+                        "Pepper do a dance",
+                    ]
+                },
+            }
+
+            msg = Message(to=to_jid)
+            msg.body = json.dumps(payload)
+            await self.send(msg)
+            print(f"Beliefs sent to {to_jid}!")
+
+            self.exit_code = "Job Finished!"
+            await self.agent.stop()
+
+    async def setup(self):
+        print("BeliefTextAgent started")
+        self.b = self.SendOnceBehaviourBlfText()
+        self.add_behaviour(self.b)
diff --git a/src/control_backend/agents/ri_command_agent.py b/src/control_backend/agents/ri_command_agent.py
index 51b8064..51e148f 100644
--- a/src/control_backend/agents/ri_command_agent.py
+++ b/src/control_backend/agents/ri_command_agent.py
@@ -4,9 +4,9 @@ import logging
 import zmq
 from spade.agent import Agent
 from spade.behaviour import CyclicBehaviour
+from zmq.asyncio import Context
 
 from control_backend.core.config import settings
-from control_backend.core.zmq_context import context
 from control_backend.schemas.ri_message import SpeechCommand
 
 logger = logging.getLogger(__name__)
@@ -56,16 +56,18 @@ class RICommandAgent(Agent):
         """
         logger.info("Setting up %s", self.jid)
 
+        context = Context.instance()
+
         # To the robot
         self.pubsocket = context.socket(zmq.PUB)
-        if self.bind:
+        if self.bind:  # TODO: Should this ever be the case?
             self.pubsocket.bind(self.address)
         else:
             self.pubsocket.connect(self.address)
 
         # Receive internal topics regarding commands
         self.subsocket = context.socket(zmq.SUB)
-        self.subsocket.connect(settings.zmq_settings.internal_comm_address)
+        self.subsocket.connect(settings.zmq_settings.internal_sub_address)
         self.subsocket.setsockopt(zmq.SUBSCRIBE, b"command")
 
         # Add behaviour to our agent
diff --git a/src/control_backend/agents/ri_communication_agent.py b/src/control_backend/agents/ri_communication_agent.py
index 0bb369d..9d9170f 100644
--- a/src/control_backend/agents/ri_communication_agent.py
+++ b/src/control_backend/agents/ri_communication_agent.py
@@ -2,21 +2,18 @@ import asyncio
 import json
 import logging
 
-import zmq
 import zmq.asyncio
 from spade.agent import Agent
 from spade.behaviour import CyclicBehaviour
+from zmq.asyncio import Context
 
 from control_backend.agents.ri_command_agent import RICommandAgent
 from control_backend.core.config import settings
-from control_backend.core.zmq_context import context
 
 logger = logging.getLogger(__name__)
 
 
 class RICommunicationAgent(Agent):
-    _pub_socket: zmq.asyncio.Socket
-    req_socket: zmq.asyncio.Socket | None
     _address = ""
     _bind = True
     connected = False
@@ -25,7 +22,6 @@ class RICommunicationAgent(Agent):
         self,
         jid: str,
         password: str,
-        pub_socket: zmq.asyncio.Socket,
         port: int = 5222,
         verify_security: bool = False,
         address="tcp://localhost:0000",
@@ -34,8 +30,8 @@ class RICommunicationAgent(Agent):
         super().__init__(jid, password, port, verify_security)
         self._address = address
         self._bind = bind
-        self.req_socket = None
-        self._pub_socket = pub_socket
+        self._req_socket: zmq.asyncio.Socket | None = None
+        self.pub_socket: zmq.asyncio.Socket | None = None
 
     class ListenBehaviour(CyclicBehaviour):
         async def run(self):
@@ -49,7 +45,7 @@ class RICommunicationAgent(Agent):
             seconds_to_wait_total = 1.0
             try:
                 await asyncio.wait_for(
-                    self.agent.req_socket.send_json(message), timeout=seconds_to_wait_total / 2
+                    self.agent._req_socket.send_json(message), timeout=seconds_to_wait_total / 2
                 )
             except TimeoutError:
                 logger.debug(
@@ -61,23 +57,13 @@ class RICommunicationAgent(Agent):
             try:
                 logger.debug(f"waiting for message for {seconds_to_wait_total / 2} seconds.")
                 message = await asyncio.wait_for(
-                    self.agent.req_socket.recv_json(), timeout=seconds_to_wait_total / 2
+                    self.agent._req_socket.recv_json(), timeout=seconds_to_wait_total / 2
                 )
 
             # We didnt get a reply :(
             except TimeoutError:
-                logger.info(
-                    f"No ping back retrieved in {seconds_to_wait_total / 2} seconds totalling"
-                    f"{seconds_to_wait_total} of time, killing myself (or maybe just laying low)."
-                )
-                # TODO: Send event to UI letting know that we've lost connection
-                topic = b"ping"
-                data = json.dumps(False).encode()
-                self.agent._pub_socket.send_multipart([topic, data])
-                await self.agent.setup()
-
-            except Exception as e:
-                logger.debug(f"Differennt exception: {e}")
+                logger.info("No ping retrieved in 3 seconds, killing myself.")
+                self.kill()
 
             logger.debug('Received message "%s"', message)
             if "endpoint" not in message:
@@ -89,46 +75,53 @@ class RICommunicationAgent(Agent):
                 case "ping":
                     topic = b"ping"
                     data = json.dumps(True).encode()
-                    await self.agent._pub_socket.send_multipart([topic, data])
+                    if self.agent.pub_socket is not None:
+                        await self.agent.pub_socket.send_multipart([topic, data])
                     await asyncio.sleep(1)
                 case _:
                     logger.info(
                         "Received message with topic different than ping, while ping expected."
                     )
 
-    async def setup_req_socket(self, force=False):
+    async def setup_sockets(self, force=False):
         """
         Sets up request socket for communication agent.
         """
-        if self.req_socket is None or force:
-            self.req_socket = context.socket(zmq.REQ)
-            if self._bind:
-                self.req_socket.bind(self._address)
+        # Bind request socket
+        if self._req_socket is None or force:
+            self._req_socket = Context.instance().socket(zmq.REQ)
+            if self._bind:  # TODO: Should this ever be the case with new architecture?
+                self._req_socket.bind(self._address)
             else:
-                self.req_socket.connect(self._address)
+                self._req_socket.connect(self._address)
 
-    async def setup(self, max_retries: int = 5):
+        # TODO: Check with Kasper
+        if self.pub_socket is None or force:
+            self.pub_socket = Context.instance().socket(zmq.PUB)
+            self.pub_socket.connect(settings.zmq_settings.internal_pub_address)
+
+    async def setup(self, max_retries: int = 100):
         """
         Try to setup the communication agent, we have 5 retries in case we dont have a response yet.
         """
         logger.info("Setting up %s", self.jid)
 
         # Bind request socket
-        await self.setup_req_socket()
+        await self.setup_sockets()
 
         retries = 0
         # Let's try a certain amount of times before failing connection
         while retries < max_retries:
             # Make sure the socket is properly setup.
-            if self.req_socket is None:
+            if self._req_socket is None:
                 continue
 
             # Send our message and receive one back:)
             message = {"endpoint": "negotiate/ports", "data": {}}
-            await self.req_socket.send_json(message)
+            await self._req_socket.send_json(message)
 
             try:
-                received_message = await asyncio.wait_for(self.req_socket.recv_json(), timeout=20.0)
+                received_message = await asyncio.wait_for(self._req_socket.recv_json(), timeout=1.0)
 
             except TimeoutError:
                 logger.warning(
@@ -173,9 +166,9 @@ class RICommunicationAgent(Agent):
                         case "main":
                             if addr != self._address:
                                 if not bind:
-                                    self.req_socket.connect(addr)
-                                else:
-                                    self.req_socket.bind(addr)
+                                    self._req_socket.connect(addr)
+                                else:  # TODO: Should this ever be the case?
+                                    self._req_socket.bind(addr)
                         case "actuation":
                             ri_commands_agent = RICommandAgent(
                                 settings.agent_settings.ri_command_agent_name
@@ -205,9 +198,17 @@ class RICommunicationAgent(Agent):
         listen_behaviour = self.ListenBehaviour()
         self.add_behaviour(listen_behaviour)
 
-        # TODO: Let UI know that we're connected >:)
+        # Let UI know that we're connected >:)
         topic = b"ping"
         data = json.dumps(True).encode()
-        await self._pub_socket.send_multipart([topic, data])
+        if self.pub_socket is None:
+            logger.error("communication agent pub socket not correctly initialized.")
+        else:
+            try:
+                await asyncio.wait_for(self.pub_socket.send_multipart([topic, data]), 5)
+            except TimeoutError:
+                logger.error(
+                    "Initial connection ping for router timed out in ri_communication_agent."
+                )
         self.connected = True
         logger.info("Finished setting up %s", self.jid)
diff --git a/src/control_backend/agents/transcription/__init__.py b/src/control_backend/agents/transcription/__init__.py
new file mode 100644
index 0000000..fd3c8c5
--- /dev/null
+++ b/src/control_backend/agents/transcription/__init__.py
@@ -0,0 +1,2 @@
+from .speech_recognizer import SpeechRecognizer as SpeechRecognizer
+from .transcription_agent import TranscriptionAgent as TranscriptionAgent
diff --git a/src/control_backend/agents/transcription/speech_recognizer.py b/src/control_backend/agents/transcription/speech_recognizer.py
new file mode 100644
index 0000000..19d82ff
--- /dev/null
+++ b/src/control_backend/agents/transcription/speech_recognizer.py
@@ -0,0 +1,108 @@
+import abc
+import sys
+
+if sys.platform == "darwin":
+    import mlx.core as mx
+    import mlx_whisper
+    from mlx_whisper.transcribe import ModelHolder
+
+import numpy as np
+import torch
+import whisper
+
+
+class SpeechRecognizer(abc.ABC):
+    def __init__(self, limit_output_length=True):
+        """
+        :param limit_output_length: When `True`, the length of the generated speech will be limited
+        by the length of the input audio and some heuristics.
+        """
+        self.limit_output_length = limit_output_length
+
+    @abc.abstractmethod
+    def load_model(self): ...
+
+    @abc.abstractmethod
+    def recognize_speech(self, audio: np.ndarray) -> str:
+        """
+        Recognize speech from the given audio sample.
+
+        :param audio: A full utterance sample. Audio must be 16 kHz, mono, np.float32, values in the
+        range [-1.0, 1.0].
+        :return: Recognized speech.
+        """
+
+    @staticmethod
+    def _estimate_max_tokens(audio: np.ndarray) -> int:
+        """
+        Estimate the maximum length of a given audio sample in tokens. Assumes a maximum speaking
+        rate of 300 words per minute (2x average), and assumes that 3 words is 4 tokens.
+
+        :param audio: The audio sample (16 kHz) to use for length estimation.
+        :return: The estimated length of the transcribed audio in tokens.
+        """
+        length_seconds = len(audio) / 16_000
+        length_minutes = length_seconds / 60
+        word_count = length_minutes * 300
+        token_count = word_count / 3 * 4
+        return int(token_count)
+
+    def _get_decode_options(self, audio: np.ndarray) -> dict:
+        """
+        :param audio: The audio sample (16 kHz) to use to determine options like max decode length.
+        :return: A dict that can be used to construct `whisper.DecodingOptions`.
+        """
+        options = {}
+        if self.limit_output_length:
+            options["sample_len"] = self._estimate_max_tokens(audio)
+        return options
+
+    @staticmethod
+    def best_type():
+        """Get the best type of SpeechRecognizer based on system capabilities."""
+        if torch.mps.is_available():
+            print("Choosing MLX Whisper model.")
+            return MLXWhisperSpeechRecognizer()
+        else:
+            print("Choosing reference Whisper model.")
+            return OpenAIWhisperSpeechRecognizer()
+
+
+class MLXWhisperSpeechRecognizer(SpeechRecognizer):
+    def __init__(self, limit_output_length=True):
+        super().__init__(limit_output_length)
+        self.was_loaded = False
+        self.model_name = "mlx-community/whisper-small.en-mlx"
+
+    def load_model(self):
+        if self.was_loaded:
+            return
+        # There appears to be no dedicated mechanism to preload a model, but this `get_model` does
+        # store it in memory for later usage
+        ModelHolder.get_model(self.model_name, mx.float16)
+        self.was_loaded = True
+
+    def recognize_speech(self, audio: np.ndarray) -> str:
+        self.load_model()
+        return mlx_whisper.transcribe(
+            audio, path_or_hf_repo=self.model_name, decode_options=self._get_decode_options(audio)
+        )["text"]
+        return mlx_whisper.transcribe(audio, path_or_hf_repo=self.model_name)["text"].strip()
+
+
+class OpenAIWhisperSpeechRecognizer(SpeechRecognizer):
+    def __init__(self, limit_output_length=True):
+        super().__init__(limit_output_length)
+        self.model = None
+
+    def load_model(self):
+        if self.model is not None:
+            return
+        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+        self.model = whisper.load_model("small.en", device=device)
+
+    def recognize_speech(self, audio: np.ndarray) -> str:
+        self.load_model()
+        return whisper.transcribe(
+            self.model, audio, decode_options=self._get_decode_options(audio)
+        )["text"]
diff --git a/src/control_backend/agents/transcription/transcription_agent.py b/src/control_backend/agents/transcription/transcription_agent.py
new file mode 100644
index 0000000..530bd68
--- /dev/null
+++ b/src/control_backend/agents/transcription/transcription_agent.py
@@ -0,0 +1,84 @@
+import asyncio
+import logging
+
+import numpy as np
+import zmq
+import zmq.asyncio as azmq
+from spade.agent import Agent
+from spade.behaviour import CyclicBehaviour
+from spade.message import Message
+
+from control_backend.agents.transcription.speech_recognizer import SpeechRecognizer
+from control_backend.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+class TranscriptionAgent(Agent):
+    """
+    An agent which listens to audio fragments with voice, transcribes them, and sends the
+    transcription to other agents.
+    """
+
+    def __init__(self, audio_in_address: str):
+        jid = settings.agent_settings.transcription_agent_name + "@" + settings.agent_settings.host
+        super().__init__(jid, settings.agent_settings.transcription_agent_name)
+
+        self.audio_in_address = audio_in_address
+        self.audio_in_socket: azmq.Socket | None = None
+
+    class Transcribing(CyclicBehaviour):
+        def __init__(self, audio_in_socket: azmq.Socket):
+            super().__init__()
+            self.audio_in_socket = audio_in_socket
+            self.speech_recognizer = SpeechRecognizer.best_type()
+            self._concurrency = asyncio.Semaphore(3)
+
+        def warmup(self):
+            """Load the transcription model into memory to speed up the first transcription."""
+            self.speech_recognizer.load_model()
+
+        async def _transcribe(self, audio: np.ndarray) -> str:
+            async with self._concurrency:
+                return await asyncio.to_thread(self.speech_recognizer.recognize_speech, audio)
+
+        async def _share_transcription(self, transcription: str):
+            """Share a transcription to the other agents that depend on it."""
+            receiver_jids = [
+                settings.agent_settings.text_belief_extractor_agent_name
+                + "@"
+                + settings.agent_settings.host,
+            ]  # Set message receivers here
+
+            for receiver_jid in receiver_jids:
+                message = Message(to=receiver_jid, body=transcription)
+                await self.send(message)
+
+        async def run(self) -> None:
+            audio = await self.audio_in_socket.recv()
+            audio = np.frombuffer(audio, dtype=np.float32)
+            speech = await self._transcribe(audio)
+            logger.info("Transcribed speech: %s", speech)
+
+            await self._share_transcription(speech)
+
+    async def stop(self):
+        self.audio_in_socket.close()
+        self.audio_in_socket = None
+        return await super().stop()
+
+    def _connect_audio_in_socket(self):
+        self.audio_in_socket = azmq.Context.instance().socket(zmq.SUB)
+        self.audio_in_socket.setsockopt_string(zmq.SUBSCRIBE, "")
+        self.audio_in_socket.connect(self.audio_in_address)
+
+    async def setup(self):
+        logger.info("Setting up %s", self.jid)
+
+        self._connect_audio_in_socket()
+
+        transcribing = self.Transcribing(self.audio_in_socket)
+        transcribing.warmup()
+        self.add_behaviour(transcribing)
+
+        logger.info("Finished setting up %s", self.jid)
diff --git a/src/control_backend/agents/vad_agent.py b/src/control_backend/agents/vad_agent.py
new file mode 100644
index 0000000..f16abf4
--- /dev/null
+++ b/src/control_backend/agents/vad_agent.py
@@ -0,0 +1,159 @@
+import logging
+
+import numpy as np
+import torch
+import zmq
+import zmq.asyncio as azmq
+from spade.agent import Agent
+from spade.behaviour import CyclicBehaviour
+
+from control_backend.agents.transcription import TranscriptionAgent
+from control_backend.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+class SocketPoller[T]:
+    """
+    Convenience class for polling a socket for data with a timeout, persisting a zmq.Poller for
+    multiple usages.
+    """
+
+    def __init__(self, socket: azmq.Socket, timeout_ms: int = 100):
+        """
+        :param socket: The socket to poll and get data from.
+        :param timeout_ms: A timeout in milliseconds to wait for data.
+        """
+        self.socket = socket
+        self.poller = zmq.Poller()
+        self.poller.register(self.socket, zmq.POLLIN)
+        self.timeout_ms = timeout_ms
+
+    async def poll(self, timeout_ms: int | None = None) -> T | None:
+        """
+        Get data from the socket, or None if the timeout is reached.
+
+        :param timeout_ms: If given, the timeout. Otherwise, `self.timeout_ms` is used.
+        :return: Data from the socket or None.
+        """
+        timeout_ms = timeout_ms or self.timeout_ms
+        socks = dict(self.poller.poll(timeout_ms))
+        if socks.get(self.socket) == zmq.POLLIN:
+            return await self.socket.recv()
+        return None
+
+
+class Streaming(CyclicBehaviour):
+    def __init__(self, audio_in_socket: azmq.Socket, audio_out_socket: azmq.Socket):
+        super().__init__()
+        self.audio_in_poller = SocketPoller[bytes](audio_in_socket)
+        self.model, _ = torch.hub.load(
+            repo_or_dir="snakers4/silero-vad", model="silero_vad", force_reload=False
+        )
+        self.audio_out_socket = audio_out_socket
+
+        self.audio_buffer = np.array([], dtype=np.float32)
+        self.i_since_speech = 100  # Used to allow small pauses in speech
+
+    async def run(self) -> None:
+        data = await self.audio_in_poller.poll()
+        if data is None:
+            if len(self.audio_buffer) > 0:
+                logger.debug("No audio data received. Discarding buffer until new data arrives.")
+                self.audio_buffer = np.array([], dtype=np.float32)
+                self.i_since_speech = 100
+            return
+
+        # copy otherwise Torch will be sad that it's immutable
+        chunk = np.frombuffer(data, dtype=np.float32).copy()
+        prob = self.model(torch.from_numpy(chunk), 16000).item()
+
+        if prob > 0.5:
+            if self.i_since_speech > 3:
+                logger.debug("Speech started.")
+            self.audio_buffer = np.append(self.audio_buffer, chunk)
+            self.i_since_speech = 0
+            return
+        self.i_since_speech += 1
+
+        # prob < 0.5, so speech maybe ended. Wait a bit more before to be more certain
+        if self.i_since_speech <= 3:
+            self.audio_buffer = np.append(self.audio_buffer, chunk)
+            return
+
+        # Speech probably ended. Make sure we have a usable amount of data.
+        if len(self.audio_buffer) >= 3 * len(chunk):
+            logger.debug("Speech ended.")
+            await self.audio_out_socket.send(self.audio_buffer[: -2 * len(chunk)].tobytes())
+
+        # At this point, we know that the speech has ended.
+        # Prepend the last chunk that had no speech, for a more fluent boundary
+        self.audio_buffer = chunk
+
+
+class VADAgent(Agent):
+    """
+    An agent which listens to an audio stream, does Voice Activity Detection (VAD), and sends
+    fragments with detected speech to other agents over ZeroMQ.
+    """
+
+    def __init__(self, audio_in_address: str, audio_in_bind: bool):
+        jid = settings.agent_settings.vad_agent_name + "@" + settings.agent_settings.host
+        super().__init__(jid, settings.agent_settings.vad_agent_name)
+
+        self.audio_in_address = audio_in_address
+        self.audio_in_bind = audio_in_bind
+
+        self.audio_in_socket: azmq.Socket | None = None
+        self.audio_out_socket: azmq.Socket | None = None
+
+    async def stop(self):
+        """
+        Stop listening to audio, stop publishing audio, close sockets.
+        """
+        if self.audio_in_socket is not None:
+            self.audio_in_socket.close()
+            self.audio_in_socket = None
+        if self.audio_out_socket is not None:
+            self.audio_out_socket.close()
+            self.audio_out_socket = None
+        return await super().stop()
+
+    def _connect_audio_in_socket(self):
+        self.audio_in_socket = azmq.Context.instance().socket(zmq.SUB)
+        self.audio_in_socket.setsockopt_string(zmq.SUBSCRIBE, "")
+        if self.audio_in_bind:
+            self.audio_in_socket.bind(self.audio_in_address)
+        else:
+            self.audio_in_socket.connect(self.audio_in_address)
+        self.audio_in_poller = SocketPoller[bytes](self.audio_in_socket)
+
+    def _connect_audio_out_socket(self) -> int | None:
+        """Returns the port bound, or None if binding failed."""
+        try:
+            self.audio_out_socket = azmq.Context.instance().socket(zmq.PUB)
+            return self.audio_out_socket.bind_to_random_port("tcp://*", max_tries=100)
+        except zmq.ZMQBindError:
+            logger.error("Failed to bind an audio output socket after 100 tries.")
+            self.audio_out_socket = None
+            return None
+
+    async def setup(self):
+        logger.info("Setting up %s", self.jid)
+
+        self._connect_audio_in_socket()
+
+        audio_out_port = self._connect_audio_out_socket()
+        if audio_out_port is None:
+            await self.stop()
+            return
+        audio_out_address = f"tcp://localhost:{audio_out_port}"
+
+        streaming = Streaming(self.audio_in_socket, self.audio_out_socket)
+        self.add_behaviour(streaming)
+
+        # Start agents dependent on the output audio fragments here
+        transcriber = TranscriptionAgent(audio_out_address)
+        await transcriber.start()
+
+        logger.info("Finished setting up %s", self.jid)
diff --git a/src/control_backend/api/v1/endpoints/message.py b/src/control_backend/api/v1/endpoints/message.py
index 1053c3c..bd88a0b 100644
--- a/src/control_backend/api/v1/endpoints/message.py
+++ b/src/control_backend/api/v1/endpoints/message.py
@@ -1,7 +1,6 @@
 import logging
 
 from fastapi import APIRouter, Request
-from zmq import Socket
 
 from control_backend.schemas.message import Message
 
@@ -17,8 +16,7 @@ async def receive_message(message: Message, request: Request):
     topic = b"message"
     body = message.model_dump_json().encode("utf-8")
 
-    pub_socket: Socket = request.app.state.internal_comm_socket
-
-    pub_socket.send_multipart([topic, body])
+    pub_socket = request.app.state.endpoints_pub_socket
+    await pub_socket.send_multipart([topic, body])
 
     return {"status": "Message received"}
diff --git a/src/control_backend/api/v1/endpoints/robot.py b/src/control_backend/api/v1/endpoints/robot.py
index e114757..7b1c2f8 100644
--- a/src/control_backend/api/v1/endpoints/robot.py
+++ b/src/control_backend/api/v1/endpoints/robot.py
@@ -5,10 +5,9 @@ import logging
 import zmq.asyncio
 from fastapi import APIRouter, Request
 from fastapi.responses import StreamingResponse
-from zmq.asyncio import Socket
+from zmq.asyncio import Context, Socket
 
 from control_backend.core.config import settings
-from control_backend.core.zmq_context import context
 from control_backend.schemas.ri_message import SpeechCommand
 
 logger = logging.getLogger(__name__)
@@ -21,6 +20,8 @@ async def receive_command(command: SpeechCommand, request: Request):
     # Validate and retrieve data.
     SpeechCommand.model_validate(command)
     topic = b"command"
+
+    # TODO: Check with Kasper
     pub_socket: Socket = request.app.state.internal_comm_socket
     pub_socket.send_multipart([topic, command.model_dump_json().encode()])
 
@@ -40,8 +41,9 @@ async def ping_stream(request: Request):
         # Set up internal socket to receive ping updates
         logger.debug("Ping stream router event stream entered.")
 
-        sub_socket = context.socket(zmq.SUB)
-        sub_socket.connect(settings.zmq_settings.internal_comm_address)
+        # TODO: Check with Kasper
+        sub_socket = Context.instance().socket(zmq.SUB)
+        sub_socket.connect(settings.zmq_settings.internal_sub_address)
         sub_socket.setsockopt(zmq.SUBSCRIBE, b"ping")
         connected = False
 
diff --git a/src/control_backend/core/config.py b/src/control_backend/core/config.py
index f48d54f..8de2403 100644
--- a/src/control_backend/core/config.py
+++ b/src/control_backend/core/config.py
@@ -3,19 +3,29 @@ from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
 class ZMQSettings(BaseModel):
-    internal_comm_address: str = "tcp://localhost:5560"
+    internal_pub_address: str = "tcp://localhost:5560"
+    internal_sub_address: str = "tcp://localhost:5561"
 
 
 class AgentSettings(BaseModel):
     host: str = "localhost"
     bdi_core_agent_name: str = "bdi_core"
     belief_collector_agent_name: str = "belief_collector"
+    text_belief_extractor_agent_name: str = "text_belief_extractor"
+    vad_agent_name: str = "vad_agent"
+    llm_agent_name: str = "llm_agent"
     test_agent_name: str = "test_agent"
+    transcription_agent_name: str = "transcription_agent"
 
     ri_communication_agent_name: str = "ri_communication_agent"
     ri_command_agent_name: str = "ri_command_agent"
 
 
+class LLMSettings(BaseModel):
+    local_llm_url: str = "http://localhost:1234/v1/chat/completions"
+    local_llm_model: str = "openai/gpt-oss-20b"
+
+
 class Settings(BaseSettings):
     app_title: str = "PepperPlus"
 
@@ -25,6 +35,8 @@ class Settings(BaseSettings):
 
     agent_settings: AgentSettings = AgentSettings()
 
+    llm_settings: LLMSettings = LLMSettings()
+
     model_config = SettingsConfigDict(env_file=".env")
 
 
diff --git a/src/control_backend/core/zmq_context.py b/src/control_backend/core/zmq_context.py
deleted file mode 100644
index a74544f..0000000
--- a/src/control_backend/core/zmq_context.py
+++ /dev/null
@@ -1,3 +0,0 @@
-from zmq.asyncio import Context
-
-context = Context()
diff --git a/src/control_backend/main.py b/src/control_backend/main.py
index a824ab1..b8e3ef3 100644
--- a/src/control_backend/main.py
+++ b/src/control_backend/main.py
@@ -3,33 +3,61 @@
 # External imports
 import contextlib
 import logging
+import threading
 
 import zmq
 from fastapi import FastAPI
 from fastapi.middleware.cors import CORSMiddleware
+from zmq.asyncio import Context
 
 from control_backend.agents.bdi.bdi_core import BDICoreAgent
+from control_backend.agents.bdi.text_extractor import TBeliefExtractor
+from control_backend.agents.belief_collector.belief_collector import BeliefCollectorAgent
+from control_backend.agents.llm.llm import LLMAgent
 
 # Internal imports
 from control_backend.agents.ri_communication_agent import RICommunicationAgent
+from control_backend.agents.vad_agent import VADAgent
 from control_backend.api.v1.router import api_router
 from control_backend.core.config import settings
-from control_backend.core.zmq_context import context
 
 logger = logging.getLogger(__name__)
 logging.basicConfig(level=logging.DEBUG)
 
 
+def setup_sockets():
+    context = Context.instance()
+
+    internal_pub_socket = context.socket(zmq.XPUB)
+    internal_pub_socket.bind(settings.zmq_settings.internal_sub_address)
+    logger.debug("Internal publishing socket bound to %s", internal_pub_socket)
+
+    internal_sub_socket = context.socket(zmq.XSUB)
+    internal_sub_socket.bind(settings.zmq_settings.internal_pub_address)
+    logger.debug("Internal subscribing socket bound to %s", internal_sub_socket)
+    try:
+        zmq.proxy(internal_sub_socket, internal_pub_socket)
+    except zmq.ZMQError:
+        logger.warning("Error while handling PUB/SUB proxy. Closing sockets.")
+    finally:
+        internal_pub_socket.close()
+        internal_sub_socket.close()
+
+
 @contextlib.asynccontextmanager
 async def lifespan(app: FastAPI):
     logger.info("%s starting up.", app.title)
 
     # Initiate sockets
-    internal_comm_socket = context.socket(zmq.PUB)
-    internal_comm_address = settings.zmq_settings.internal_comm_address
-    internal_comm_socket.bind(internal_comm_address)
-    app.state.internal_comm_socket = internal_comm_socket
-    logger.info("Internal publishing socket bound to %s", internal_comm_socket)
+    proxy_thread = threading.Thread(target=setup_sockets)
+    proxy_thread.daemon = True
+    proxy_thread.start()
+
+    context = Context.instance()
+
+    endpoints_pub_socket = context.socket(zmq.PUB)
+    endpoints_pub_socket.connect(settings.zmq_settings.internal_pub_address)
+    app.state.endpoints_pub_socket = endpoints_pub_socket
 
     # Initiate agents
     ri_communication_agent = RICommunicationAgent(
@@ -37,12 +65,17 @@ async def lifespan(app: FastAPI):
         + "@"
         + settings.agent_settings.host,
         password=settings.agent_settings.ri_communication_agent_name,
-        pub_socket=internal_comm_socket,
         address="tcp://*:5555",
         bind=True,
     )
     await ri_communication_agent.start()
 
+    llm_agent = LLMAgent(
+        settings.agent_settings.llm_agent_name + "@" + settings.agent_settings.host,
+        settings.agent_settings.llm_agent_name,
+    )
+    await llm_agent.start()
+
     bdi_core = BDICoreAgent(
         settings.agent_settings.bdi_core_agent_name + "@" + settings.agent_settings.host,
         settings.agent_settings.bdi_core_agent_name,
@@ -50,6 +83,23 @@ async def lifespan(app: FastAPI):
     )
     await bdi_core.start()
 
+    belief_collector = BeliefCollectorAgent(
+        settings.agent_settings.belief_collector_agent_name + "@" + settings.agent_settings.host,
+        settings.agent_settings.belief_collector_agent_name,
+    )
+    await belief_collector.start()
+
+    text_belief_extractor = TBeliefExtractor(
+        settings.agent_settings.text_belief_extractor_agent_name
+        + "@"
+        + settings.agent_settings.host,
+        settings.agent_settings.text_belief_extractor_agent_name,
+    )
+    await text_belief_extractor.start()
+
+    _temp_vad_agent = VADAgent("tcp://localhost:5558", False)
+    await _temp_vad_agent.start()
+
     yield
 
     logger.info("%s shutting down.", app.title)
diff --git a/test/integration/agents/test_ri_commands_agent.py b/test/integration/agents/test_ri_commands_agent.py
index 4249401..a4902b5 100644
--- a/test/integration/agents/test_ri_commands_agent.py
+++ b/test/integration/agents/test_ri_commands_agent.py
@@ -11,25 +11,27 @@ from control_backend.agents.ri_command_agent import RICommandAgent
 async def test_setup_bind(monkeypatch):
     """Test setup with bind=True"""
     fake_socket = MagicMock()
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
+
+    # Patch Context.instance() to return fake_context
     monkeypatch.setattr(
-        "control_backend.agents.ri_command_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_command_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     agent = RICommandAgent("test@server", "password", address="tcp://localhost:5555", bind=True)
+
     monkeypatch.setattr(
         "control_backend.agents.ri_command_agent.settings",
-        MagicMock(zmq_settings=MagicMock(internal_comm_address="tcp://internal:1234")),
+        MagicMock(zmq_settings=MagicMock(internal_sub_address="tcp://internal:1234")),
     )
 
     await agent.setup()
 
-    # Ensure PUB socket bound
     fake_socket.bind.assert_any_call("tcp://localhost:5555")
-    # Ensure SUB socket connected to internal address and subscribed
     fake_socket.connect.assert_any_call("tcp://internal:1234")
     fake_socket.setsockopt.assert_any_call(zmq.SUBSCRIBE, b"command")
-
-    # Ensure behaviour attached
     assert any(isinstance(b, agent.SendCommandsBehaviour) for b in agent.behaviours)
 
 
@@ -37,19 +39,23 @@ async def test_setup_bind(monkeypatch):
 async def test_setup_connect(monkeypatch):
     """Test setup with bind=False"""
     fake_socket = MagicMock()
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
+
+    # Patch Context.instance() to return fake_context
     monkeypatch.setattr(
-        "control_backend.agents.ri_command_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_command_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     agent = RICommandAgent("test@server", "password", address="tcp://localhost:5555", bind=False)
     monkeypatch.setattr(
         "control_backend.agents.ri_command_agent.settings",
-        MagicMock(zmq_settings=MagicMock(internal_comm_address="tcp://internal:1234")),
+        MagicMock(zmq_settings=MagicMock(internal_sub_address="tcp://internal:1234")),
     )
 
     await agent.setup()
 
-    # Ensure PUB socket connected
     fake_socket.connect.assert_any_call("tcp://localhost:5555")
 
 
diff --git a/test/integration/agents/test_ri_communication_agent.py b/test/integration/agents/test_ri_communication_agent.py
index baeb717..9febf20 100644
--- a/test/integration/agents/test_ri_communication_agent.py
+++ b/test/integration/agents/test_ri_communication_agent.py
@@ -93,12 +93,14 @@ async def test_setup_creates_socket_and_negotiate_1(monkeypatch):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_correct_negototiate_1()
-
-    fake_pub_socket = AsyncMock()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -107,13 +109,11 @@ async def test_setup_creates_socket_and_negotiate_1(monkeypatch):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
 
         # --- Act ---
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=False,
         )
@@ -143,10 +143,14 @@ async def test_setup_creates_socket_and_negotiate_2(monkeypatch):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_correct_negototiate_2()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -155,13 +159,11 @@ async def test_setup_creates_socket_and_negotiate_2(monkeypatch):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
 
         # --- Act ---
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=False,
         )
@@ -191,10 +193,14 @@ async def test_setup_creates_socket_and_negotiate_3(monkeypatch, caplog):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_wrong_negototiate_1()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -206,13 +212,11 @@ async def test_setup_creates_socket_and_negotiate_3(monkeypatch, caplog):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
         # --- Act ---
         with caplog.at_level("ERROR"):
             agent = RICommunicationAgent(
                 "test@server",
                 "password",
-                pub_socket=fake_pub_socket,
                 address="tcp://localhost:5555",
                 bind=False,
             )
@@ -239,10 +243,14 @@ async def test_setup_creates_socket_and_negotiate_4(monkeypatch):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_correct_negototiate_3()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -251,12 +259,10 @@ async def test_setup_creates_socket_and_negotiate_4(monkeypatch):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
         # --- Act ---
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=True,
         )
@@ -286,10 +292,14 @@ async def test_setup_creates_socket_and_negotiate_5(monkeypatch):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_correct_negototiate_4()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -298,12 +308,10 @@ async def test_setup_creates_socket_and_negotiate_5(monkeypatch):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
         # --- Act ---
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=False,
         )
@@ -333,10 +341,14 @@ async def test_setup_creates_socket_and_negotiate_6(monkeypatch):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_correct_negototiate_5()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -345,12 +357,10 @@ async def test_setup_creates_socket_and_negotiate_6(monkeypatch):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
         # --- Act ---
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=False,
         )
@@ -380,10 +390,14 @@ async def test_setup_creates_socket_and_negotiate_7(monkeypatch, caplog):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = fake_json_invalid_id_negototiate()
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Mock RICommandAgent agent startup
@@ -395,14 +409,12 @@ async def test_setup_creates_socket_and_negotiate_7(monkeypatch, caplog):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
 
         # --- Act ---
         with caplog.at_level("WARNING"):
             agent = RICommunicationAgent(
                 "test@server",
                 "password",
-                pub_socket=fake_pub_socket,
                 address="tcp://localhost:5555",
                 bind=False,
             )
@@ -426,10 +438,14 @@ async def test_setup_creates_socket_and_negotiate_timeout(monkeypatch, caplog):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = AsyncMock(side_effect=asyncio.TimeoutError)
+    fake_socket.send_multipart = AsyncMock()
 
     # Mock context.socket to return our fake socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     with patch(
@@ -437,14 +453,12 @@ async def test_setup_creates_socket_and_negotiate_timeout(monkeypatch, caplog):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
 
         # --- Act ---
         with caplog.at_level("WARNING"):
             agent = RICommunicationAgent(
                 "test@server",
                 "password",
-                pub_socket=fake_pub_socket,
                 address="tcp://localhost:5555",
                 bind=False,
             )
@@ -466,11 +480,11 @@ async def test_listen_behaviour_ping_correct(caplog):
     fake_socket = AsyncMock()
     fake_socket.send_json = AsyncMock()
     fake_socket.recv_json = AsyncMock(return_value={"endpoint": "ping", "data": {}})
-    fake_pub_socket = AsyncMock()
+    fake_socket.send_multipart = AsyncMock()
 
     # TODO: Integration test between actual server and password needed for spade agents
-    agent = RICommunicationAgent("test@server", "password", fake_pub_socket)
-    agent.req_socket = fake_socket
+    agent = RICommunicationAgent("test@server", "password")
+    agent._req_socket = fake_socket
 
     behaviour = agent.ListenBehaviour()
     agent.add_behaviour(behaviour)
@@ -505,7 +519,7 @@ async def test_listen_behaviour_ping_wrong_endpoint(caplog):
     fake_pub_socket = AsyncMock()
 
     agent = RICommunicationAgent("test@server", "password", fake_pub_socket)
-    agent.req_socket = fake_socket
+    agent._req_socket = fake_socket
 
     behaviour = agent.ListenBehaviour()
     agent.add_behaviour(behaviour)
@@ -525,10 +539,10 @@ async def test_listen_behaviour_timeout(caplog):
     fake_socket.send_json = AsyncMock()
     # recv_json will never resolve, simulate timeout
     fake_socket.recv_json = AsyncMock(side_effect=asyncio.TimeoutError)
-    fake_pub_socket = AsyncMock()
+    fake_socket.send_multipart = AsyncMock()
 
-    agent = RICommunicationAgent("test@server", "password", fake_pub_socket)
-    agent.req_socket = fake_socket
+    agent = RICommunicationAgent("test@server", "password")
+    agent._req_socket = fake_socket
 
     behaviour = agent.ListenBehaviour()
     agent.add_behaviour(behaviour)
@@ -546,6 +560,7 @@ async def test_listen_behaviour_ping_no_endpoint(caplog):
     """
     fake_socket = AsyncMock()
     fake_socket.send_json = AsyncMock()
+    fake_socket.send_multipart = AsyncMock()
 
     # This is a message without endpoint >:(
     fake_socket.recv_json = AsyncMock(
@@ -553,10 +568,9 @@ async def test_listen_behaviour_ping_no_endpoint(caplog):
             "data": "I dont have an endpoint >:)",
         }
     )
-    fake_pub_socket = AsyncMock()
 
-    agent = RICommunicationAgent("test@server", "password", fake_pub_socket)
-    agent.req_socket = fake_socket
+    agent = RICommunicationAgent("test@server", "password")
+    agent._req_socket = fake_socket
 
     behaviour = agent.ListenBehaviour()
     agent.add_behaviour(behaviour)
@@ -574,18 +588,20 @@ async def test_listen_behaviour_ping_no_endpoint(caplog):
 async def test_setup_unexpected_exception(monkeypatch, caplog):
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
-    fake_pub_socket = AsyncMock()
     # Simulate unexpected exception during recv_json()
     fake_socket.recv_json = AsyncMock(side_effect=Exception("boom!"))
+    fake_socket.send_multipart = AsyncMock()
 
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     agent = RICommunicationAgent(
         "test@server",
         "password",
-        pub_socket=fake_pub_socket,
         address="tcp://localhost:5555",
         bind=False,
     )
@@ -602,6 +618,7 @@ async def test_setup_unpacking_exception(monkeypatch, caplog):
     # --- Arrange ---
     fake_socket = MagicMock()
     fake_socket.send_json = AsyncMock()
+    fake_socket.send_multipart = AsyncMock()
 
     # Make recv_json return malformed negotiation data to trigger unpacking exception
     malformed_data = {
@@ -611,8 +628,11 @@ async def test_setup_unpacking_exception(monkeypatch, caplog):
     fake_socket.recv_json = AsyncMock(return_value=malformed_data)
 
     # Patch context.socket
+    fake_context = MagicMock()
+    fake_context.socket.return_value = fake_socket
     monkeypatch.setattr(
-        "control_backend.agents.ri_communication_agent.context.socket", lambda _: fake_socket
+        "control_backend.agents.ri_communication_agent.Context",
+        MagicMock(instance=MagicMock(return_value=fake_context)),
     )
 
     # Patch RICommandAgent so it won't actually start
@@ -621,12 +641,10 @@ async def test_setup_unpacking_exception(monkeypatch, caplog):
     ) as MockCommandAgent:
         fake_agent_instance = MockCommandAgent.return_value
         fake_agent_instance.start = AsyncMock()
-        fake_pub_socket = AsyncMock()
 
         agent = RICommunicationAgent(
             "test@server",
             "password",
-            pub_socket=fake_pub_socket,
             address="tcp://localhost:5555",
             bind=False,
         )
diff --git a/test/integration/agents/vad_agent/speech_with_pauses_16k_1c_float32.wav b/test/integration/agents/vad_agent/speech_with_pauses_16k_1c_float32.wav
new file mode 100644
index 0000000..530bc0a
Binary files /dev/null and b/test/integration/agents/vad_agent/speech_with_pauses_16k_1c_float32.wav differ
diff --git a/test/integration/agents/vad_agent/test_vad_agent.py b/test/integration/agents/vad_agent/test_vad_agent.py
new file mode 100644
index 0000000..54c9d82
--- /dev/null
+++ b/test/integration/agents/vad_agent/test_vad_agent.py
@@ -0,0 +1,108 @@
+import random
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+import zmq
+from spade.agent import Agent
+
+from control_backend.agents.vad_agent import VADAgent
+
+
+@pytest.fixture
+def zmq_context(mocker):
+    return mocker.patch("control_backend.agents.vad_agent.zmq_context")
+
+
+@pytest.fixture
+def streaming(mocker):
+    return mocker.patch("control_backend.agents.vad_agent.Streaming")
+
+
+@pytest.fixture
+def transcription_agent(mocker):
+    return mocker.patch("control_backend.agents.vad_agent.TranscriptionAgent", autospec=True)
+
+
+@pytest.mark.asyncio
+async def test_normal_setup(streaming, transcription_agent):
+    """
+    Test that during normal setup, the VAD agent creates a Streaming behavior and creates audio
+    sockets, and starts the TranscriptionAgent without loading real models.
+    """
+    vad_agent = VADAgent("tcp://localhost:12345", False)
+    vad_agent.add_behaviour = MagicMock()
+
+    await vad_agent.setup()
+
+    streaming.assert_called_once()
+    vad_agent.add_behaviour.assert_called_once_with(streaming.return_value)
+    transcription_agent.assert_called_once()
+    transcription_agent.return_value.start.assert_called_once()
+    assert vad_agent.audio_in_socket is not None
+    assert vad_agent.audio_out_socket is not None
+
+
+@pytest.mark.parametrize("do_bind", [True, False])
+def test_in_socket_creation(zmq_context, do_bind: bool):
+    """
+    Test that the VAD agent creates an audio input socket, differentiating between binding and
+    connecting.
+    """
+    vad_agent = VADAgent(f"tcp://{'*' if do_bind else 'localhost'}:12345", do_bind)
+
+    vad_agent._connect_audio_in_socket()
+
+    assert vad_agent.audio_in_socket is not None
+
+    zmq_context.socket.assert_called_once_with(zmq.SUB)
+    zmq_context.socket.return_value.setsockopt_string.assert_called_once_with(zmq.SUBSCRIBE, "")
+
+    if do_bind:
+        zmq_context.socket.return_value.bind.assert_called_once_with("tcp://*:12345")
+    else:
+        zmq_context.socket.return_value.connect.assert_called_once_with("tcp://localhost:12345")
+
+
+def test_out_socket_creation(zmq_context):
+    """
+    Test that the VAD agent creates an audio output socket correctly.
+    """
+    vad_agent = VADAgent("tcp://localhost:12345", False)
+
+    vad_agent._connect_audio_out_socket()
+
+    assert vad_agent.audio_out_socket is not None
+
+    zmq_context.socket.assert_called_once_with(zmq.PUB)
+    zmq_context.socket.return_value.bind_to_random_port.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_out_socket_creation_failure(zmq_context):
+    """
+    Test setup failure when the audio output socket cannot be created.
+    """
+    with patch.object(Agent, "stop", new_callable=AsyncMock) as mock_super_stop:
+        zmq_context.socket.return_value.bind_to_random_port.side_effect = zmq.ZMQBindError
+        vad_agent = VADAgent("tcp://localhost:12345", False)
+
+        await vad_agent.setup()
+
+        assert vad_agent.audio_out_socket is None
+        mock_super_stop.assert_called_once()
+
+
+@pytest.mark.asyncio
+async def test_stop(zmq_context, transcription_agent):
+    """
+    Test that when the VAD agent is stopped, the sockets are closed correctly.
+    """
+    vad_agent = VADAgent("tcp://localhost:12345", False)
+    zmq_context.socket.return_value.bind_to_random_port.return_value = random.randint(1000, 10000)
+
+    await vad_agent.setup()
+    await vad_agent.stop()
+
+    assert zmq_context.socket.return_value.close.call_count == 2
+    assert vad_agent.audio_in_socket is None
+    assert vad_agent.audio_out_socket is None
diff --git a/test/integration/agents/vad_agent/test_vad_with_audio.py b/test/integration/agents/vad_agent/test_vad_with_audio.py
new file mode 100644
index 0000000..7d10aa3
--- /dev/null
+++ b/test/integration/agents/vad_agent/test_vad_with_audio.py
@@ -0,0 +1,57 @@
+import os
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+import soundfile as sf
+import zmq
+
+from control_backend.agents.vad_agent import Streaming
+
+
+def get_audio_chunks() -> list[bytes]:
+    curr_file = os.path.realpath(__file__)
+    curr_dir = os.path.dirname(curr_file)
+    file = f"{curr_dir}/speech_with_pauses_16k_1c_float32.wav"
+
+    chunk_size = 512
+
+    chunks = []
+
+    with sf.SoundFile(file, "r") as f:
+        assert f.samplerate == 16000
+        assert f.channels == 1
+        assert f.subtype == "FLOAT"
+
+        while True:
+            data = f.read(chunk_size, dtype="float32")
+            if len(data) != chunk_size:
+                break
+
+            chunks.append(data.tobytes())
+
+    return chunks
+
+
+@pytest.mark.asyncio
+async def test_real_audio(mocker):
+    """
+    Test the VAD agent with only input and output mocked. Using the real model, using real audio as
+    input. Ensure that it outputs some fragments with audio.
+    """
+    audio_chunks = get_audio_chunks()
+    audio_in_socket = AsyncMock()
+    audio_in_socket.recv.side_effect = audio_chunks
+
+    mock_poller: MagicMock = mocker.patch("control_backend.agents.vad_agent.zmq.Poller")
+    mock_poller.return_value.poll.return_value = [(audio_in_socket, zmq.POLLIN)]
+
+    audio_out_socket = AsyncMock()
+
+    vad_streamer = Streaming(audio_in_socket, audio_out_socket)
+    for _ in audio_chunks:
+        await vad_streamer.run()
+
+    audio_out_socket.send.assert_called()
+    for args in audio_out_socket.send.call_args_list:
+        assert isinstance(args[0][0], bytes)
+        assert len(args[0][0]) >= 512 * 4 * 3  # Should be at least 3 chunks of audio
diff --git a/test/unit/agents/bdi/behaviours/test_belief_setter.py b/test/unit/agents/bdi/behaviours/test_belief_setter.py
index b8f5570..c7bb0e9 100644
--- a/test/unit/agents/bdi/behaviours/test_belief_setter.py
+++ b/test/unit/agents/bdi/behaviours/test_belief_setter.py
@@ -4,7 +4,7 @@ from unittest.mock import AsyncMock, MagicMock, call
 
 import pytest
 
-from control_backend.agents.bdi.behaviours.belief_setter import BeliefSetter
+from control_backend.agents.bdi.behaviours.belief_setter import BeliefSetterBehaviour
 
 # Define a constant for the collector agent name to use in tests
 COLLECTOR_AGENT_NAME = "belief_collector"
@@ -22,16 +22,14 @@ def mock_agent(mocker):
 
 @pytest.fixture
 def belief_setter(mock_agent, mocker):
-    """Fixture to create an instance of BeliefSetter with a mocked agent."""
+    """Fixture to create an instance of BeliefSetterBehaviour with a mocked agent."""
     # Patch the settings to use a predictable agent name
     mocker.patch(
         "control_backend.agents.bdi.behaviours.belief_setter.settings.agent_settings.belief_collector_agent_name",
         COLLECTOR_AGENT_NAME,
     )
-    # Patch asyncio.sleep to prevent tests from actually waiting
-    mocker.patch("asyncio.sleep", return_value=None)
 
-    setter = BeliefSetter()
+    setter = BeliefSetterBehaviour()
     setter.agent = mock_agent
     # Mock the receive method, we will control its return value in each test
     setter.receive = AsyncMock()
@@ -115,7 +113,7 @@ def test_process_belief_message_valid_json(belief_setter, mocker):
     Test processing a valid belief message with correct thread and JSON body.
     """
     # Arrange
-    beliefs_payload = {"is_hot": [["kitchen"]], "is_clean": [["kitchen"], ["bathroom"]]}
+    beliefs_payload = {"is_hot": ["kitchen"], "is_clean": ["kitchen", "bathroom"]}
     msg = create_mock_message(
         sender_node=COLLECTOR_AGENT_JID, body=json.dumps(beliefs_payload), thread="beliefs"
     )
@@ -185,8 +183,8 @@ def test_set_beliefs_success(belief_setter, mock_agent, caplog):
     """
     # Arrange
     beliefs_to_set = {
-        "is_hot": [["kitchen"], ["living_room"]],
-        "door_is": [["front_door", "closed"]],
+        "is_hot": ["kitchen"],
+        "door_opened": ["front_door", "back_door"],
     }
 
     # Act
@@ -196,29 +194,38 @@ def test_set_beliefs_success(belief_setter, mock_agent, caplog):
     # Assert
     expected_calls = [
         call("is_hot", "kitchen"),
-        call("is_hot", "living_room"),
-        call("door_is", "front_door", "closed"),
+        call("door_opened", "front_door", "back_door"),
     ]
     mock_agent.bdi.set_belief.assert_has_calls(expected_calls, any_order=True)
-    assert mock_agent.bdi.set_belief.call_count == 3
+    assert mock_agent.bdi.set_belief.call_count == 2
 
     # Check logs
     assert "Set belief is_hot with arguments ['kitchen']" in caplog.text
-    assert "Set belief is_hot with arguments ['living_room']" in caplog.text
-    assert "Set belief door_is with arguments ['front_door', 'closed']" in caplog.text
+    assert "Set belief door_opened with arguments ['front_door', 'back_door']" in caplog.text
 
 
-def test_set_beliefs_bdi_not_initialized(belief_setter, mock_agent, caplog):
-    """
-    Test that a warning is logged if the agent's BDI is not initialized.
-    """
-    # Arrange
-    mock_agent.bdi = None  # Simulate BDI not being ready
-    beliefs_to_set = {"is_hot": [["kitchen"]]}
+# def test_responded_unset(belief_setter, mock_agent):
+#     # Arrange
+#     new_beliefs = {"user_said": ["message"]}
+#
+#     # Act
+#     belief_setter._set_beliefs(new_beliefs)
+#
+#     # Assert
+#     mock_agent.bdi.set_belief.assert_has_calls([call("user_said", "message")])
+#     mock_agent.bdi.remove_belief.assert_has_calls([call("responded")])
 
-    # Act
-    with caplog.at_level(logging.WARNING):
-        belief_setter._set_beliefs(beliefs_to_set)
-
-    # Assert
-    assert "Cannot set beliefs, since agent's BDI is not yet initialized." in caplog.text
+# def test_set_beliefs_bdi_not_initialized(belief_setter, mock_agent, caplog):
+#     """
+#     Test that a warning is logged if the agent's BDI is not initialized.
+#     """
+#     # Arrange
+#     mock_agent.bdi = None  # Simulate BDI not being ready
+#     beliefs_to_set = {"is_hot": ["kitchen"]}
+#
+#     # Act
+#     with caplog.at_level(logging.WARNING):
+#         belief_setter._set_beliefs(beliefs_to_set)
+#
+#     # Assert
+#     assert "Cannot set beliefs, since agent's BDI is not yet initialized." in caplog.text
diff --git a/test/unit/agents/belief_collector/behaviours/test_continuous_collect.py b/test/unit/agents/belief_collector/behaviours/test_continuous_collect.py
new file mode 100644
index 0000000..79957f0
--- /dev/null
+++ b/test/unit/agents/belief_collector/behaviours/test_continuous_collect.py
@@ -0,0 +1,242 @@
+import json
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from control_backend.agents.belief_collector.behaviours.continuous_collect import (
+    ContinuousBeliefCollector,
+)
+
+
+@pytest.fixture
+def mock_agent(mocker):
+    """Fixture to create a mock Agent."""
+    agent = MagicMock()
+    agent.jid = "belief_collector_agent@test"
+    return agent
+
+
+@pytest.fixture
+def continuous_collector(mock_agent, mocker):
+    """Fixture to create an instance of ContinuousBeliefCollector with a mocked agent."""
+    # Patch asyncio.sleep to prevent tests from actually waiting
+    mocker.patch("asyncio.sleep", return_value=None)
+
+    collector = ContinuousBeliefCollector()
+    collector.agent = mock_agent
+    # Mock the receive method, we will control its return value in each test
+    collector.receive = AsyncMock()
+    return collector
+
+
+@pytest.mark.asyncio
+async def test_run_no_message_received(continuous_collector, mocker):
+    """
+    Test that when no message is received, _process_message is not called.
+    """
+    # Arrange
+    continuous_collector.receive.return_value = None
+    mocker.patch.object(continuous_collector, "_process_message")
+
+    # Act
+    await continuous_collector.run()
+
+    # Assert
+    continuous_collector._process_message.assert_not_called()
+
+
+@pytest.mark.asyncio
+async def test_run_message_received(continuous_collector, mocker):
+    """
+    Test that when a message is received, _process_message is called with that message.
+    """
+    # Arrange
+    mock_msg = MagicMock()
+    continuous_collector.receive.return_value = mock_msg
+    mocker.patch.object(continuous_collector, "_process_message")
+
+    # Act
+    await continuous_collector.run()
+
+    # Assert
+    continuous_collector._process_message.assert_awaited_once_with(mock_msg)
+
+
+@pytest.mark.asyncio
+async def test_process_message_invalid(continuous_collector, mocker):
+    """
+    Test that when an invalid JSON message is received, a warning is logged and processing stops.
+    """
+    # Arrange
+    invalid_json = "this is not json"
+    msg = MagicMock()
+    msg.body = invalid_json
+    msg.sender = "belief_text_agent_mock@test"
+
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+
+    # Act
+    await continuous_collector._process_message(msg)
+
+    # Assert
+    logger_mock.warning.assert_called_once()
+
+
+def test_get_sender_from_message(continuous_collector):
+    """
+    Test that _sender_node correctly extracts the sender node from the message JID.
+    """
+    # Arrange
+    msg = MagicMock()
+    msg.sender = "agent_node@host/resource"
+
+    # Act
+    sender_node = continuous_collector._sender_node(msg)
+
+    # Assert
+    assert sender_node == "agent_node"
+
+
+@pytest.mark.asyncio
+async def test_routes_to_handle_belief_text_by_type(continuous_collector, mocker):
+    msg = MagicMock()
+    msg.body = json.dumps({"type": "belief_extraction_text", "beliefs": {"user_said": [["hi"]]}})
+    msg.sender = "anyone@test"
+    spy = mocker.patch.object(continuous_collector, "_handle_belief_text", new=AsyncMock())
+    await continuous_collector._process_message(msg)
+    spy.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_routes_to_handle_belief_text_by_sender(continuous_collector, mocker):
+    msg = MagicMock()
+    msg.body = json.dumps({"beliefs": {"user_said": [["hi"]]}})  # no type
+    msg.sender = "belief_text_agent_mock@test"
+    spy = mocker.patch.object(continuous_collector, "_handle_belief_text", new=AsyncMock())
+    await continuous_collector._process_message(msg)
+    spy.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_routes_to_handle_emo_text(continuous_collector, mocker):
+    msg = MagicMock()
+    msg.body = json.dumps({"type": "emotion_extraction_text"})
+    msg.sender = "anyone@test"
+    spy = mocker.patch.object(continuous_collector, "_handle_emo_text", new=AsyncMock())
+    await continuous_collector._process_message(msg)
+    spy.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_unrecognized_message_logs_info(continuous_collector, mocker):
+    msg = MagicMock()
+    msg.body = json.dumps({"type": "something_else"})
+    msg.sender = "x@test"
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+    await continuous_collector._process_message(msg)
+    logger_mock.info.assert_any_call(
+        "BeliefCollector: unrecognized message (sender=%s, type=%r). Ignoring.",
+        "x",
+        "something_else",
+    )
+
+
+@pytest.mark.asyncio
+async def test_belief_text_no_beliefs(continuous_collector, mocker):
+    msg_payload = {"type": "belief_extraction_text"}  # no 'beliefs'
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+    await continuous_collector._handle_belief_text(msg_payload, "origin_node")
+    logger_mock.info.assert_any_call("BeliefCollector: no beliefs to process.")
+
+
+@pytest.mark.asyncio
+async def test_belief_text_beliefs_not_dict(continuous_collector, mocker):
+    payload = {"type": "belief_extraction_text", "beliefs": ["not", "a", "dict"]}
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+    await continuous_collector._handle_belief_text(payload, "origin")
+    logger_mock.warning.assert_any_call(
+        "BeliefCollector: 'beliefs' is not a dict: %r", ["not", "a", "dict"]
+    )
+
+
+@pytest.mark.asyncio
+async def test_belief_text_values_not_lists(continuous_collector, mocker):
+    payload = {"type": "belief_extraction_text", "beliefs": {"user_said": "not-a-list"}}
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+    await continuous_collector._handle_belief_text(payload, "origin")
+    logger_mock.warning.assert_any_call(
+        "BeliefCollector: 'beliefs' values are not all lists: %r", {"user_said": "not-a-list"}
+    )
+
+
+@pytest.mark.asyncio
+async def test_belief_text_happy_path_logs_items_and_sends(continuous_collector, mocker):
+    payload = {"type": "belief_extraction_text", "beliefs": {"user_said": ["hello test", "No"]}}
+    # Your code calls self.send(..); patch it
+    # (or switch implementation to self.agent.send and patch that)
+    continuous_collector.send = AsyncMock()
+    logger_mock = mocker.patch(
+        "control_backend.agents.belief_collector.behaviours.continuous_collect.logger"
+    )
+    await continuous_collector._handle_belief_text(payload, "belief_text_agent_mock")
+
+    logger_mock.info.assert_any_call("BeliefCollector: forwarding %d beliefs.", 1)
+    # and the item logs:
+    logger_mock.info.assert_any_call("  - %s %s", "user_said", "hello test")
+    logger_mock.info.assert_any_call("  - %s %s", "user_said", "No")
+    # make sure we attempted a send
+    continuous_collector.send.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_send_beliefs_noop_on_empty(continuous_collector):
+    continuous_collector.send = AsyncMock()
+    await continuous_collector._send_beliefs_to_bdi([], origin="o")
+    continuous_collector.send.assert_not_awaited()
+
+
+# @pytest.mark.asyncio
+# async def test_send_beliefs_sends_json_packet(continuous_collector):
+#     # Patch .send and capture the message body
+#     sent = {}
+#
+#     async def _fake_send(msg):
+#         sent["body"] = msg.body
+#         sent["to"] = str(msg.to)
+#
+#     continuous_collector.send = AsyncMock(side_effect=_fake_send)
+#     beliefs = ["user_said hello", "user_said No"]
+#     await continuous_collector._send_beliefs_to_bdi(beliefs, origin="origin_node")
+#
+#     assert "belief_packet" in json.loads(sent["body"])["type"]
+#     assert json.loads(sent["body"])["beliefs"] == beliefs
+
+
+def test_sender_node_no_sender_returns_literal(continuous_collector):
+    msg = MagicMock()
+    msg.sender = None
+    assert continuous_collector._sender_node(msg) == "no_sender"
+
+
+def test_sender_node_without_at(continuous_collector):
+    msg = MagicMock()
+    msg.sender = "localpartonly"
+    assert continuous_collector._sender_node(msg) == "localpartonly"
+
+
+@pytest.mark.asyncio
+async def test_belief_text_coerces_non_strings(continuous_collector, mocker):
+    payload = {"type": "belief_extraction_text", "beliefs": {"user_said": [["hi", 123]]}}
+    continuous_collector.send = AsyncMock()
+    await continuous_collector._handle_belief_text(payload, "origin")
+    continuous_collector.send.assert_awaited_once()
diff --git a/test/unit/agents/test_vad_socket_poller.py b/test/unit/agents/test_vad_socket_poller.py
new file mode 100644
index 0000000..aaf8d0f
--- /dev/null
+++ b/test/unit/agents/test_vad_socket_poller.py
@@ -0,0 +1,46 @@
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+import zmq
+
+from control_backend.agents.vad_agent import SocketPoller
+
+
+@pytest.fixture
+def socket():
+    return AsyncMock()
+
+
+@pytest.mark.asyncio
+async def test_socket_poller_with_data(socket, mocker):
+    socket_data = b"test"
+    socket.recv.return_value = socket_data
+
+    mock_poller: MagicMock = mocker.patch("control_backend.agents.vad_agent.zmq.Poller")
+    mock_poller.return_value.poll.return_value = [(socket, zmq.POLLIN)]
+
+    poller = SocketPoller(socket)
+    # Calling `poll` twice to be able to check that the poller is reused
+    await poller.poll()
+    data = await poller.poll()
+
+    assert data == socket_data
+
+    # Ensure that the poller was reused
+    mock_poller.assert_called_once_with()
+    mock_poller.return_value.register.assert_called_once_with(socket, zmq.POLLIN)
+
+    assert socket.recv.call_count == 2
+
+
+@pytest.mark.asyncio
+async def test_socket_poller_no_data(socket, mocker):
+    mock_poller: MagicMock = mocker.patch("control_backend.agents.vad_agent.zmq.Poller")
+    mock_poller.return_value.poll.return_value = []
+
+    poller = SocketPoller(socket)
+    data = await poller.poll()
+
+    assert data is None
+
+    socket.recv.assert_not_called()
diff --git a/test/unit/agents/test_vad_streaming.py b/test/unit/agents/test_vad_streaming.py
new file mode 100644
index 0000000..9b38cd0
--- /dev/null
+++ b/test/unit/agents/test_vad_streaming.py
@@ -0,0 +1,95 @@
+from unittest.mock import AsyncMock, MagicMock
+
+import numpy as np
+import pytest
+
+from control_backend.agents.vad_agent import Streaming
+
+
+@pytest.fixture
+def audio_in_socket():
+    return AsyncMock()
+
+
+@pytest.fixture
+def audio_out_socket():
+    return AsyncMock()
+
+
+@pytest.fixture
+def streaming(audio_in_socket, audio_out_socket):
+    import torch
+
+    torch.hub.load.return_value = (..., ...)  # Mock
+    return Streaming(audio_in_socket, audio_out_socket)
+
+
+async def simulate_streaming_with_probabilities(streaming, probabilities: list[float]):
+    """
+    Simulates a streaming scenario with given VAD model probabilities for testing purposes.
+
+    :param streaming: The streaming component to be tested.
+    :param probabilities: A list of probabilities representing the outputs of the VAD model.
+    """
+    model_item = MagicMock()
+    model_item.item.side_effect = probabilities
+    streaming.model = MagicMock()
+    streaming.model.return_value = model_item
+
+    audio_in_poller = AsyncMock()
+    audio_in_poller.poll.return_value = np.empty(shape=512, dtype=np.float32)
+    streaming.audio_in_poller = audio_in_poller
+
+    for _ in probabilities:
+        await streaming.run()
+
+
+@pytest.mark.asyncio
+async def test_voice_activity_detected(audio_in_socket, audio_out_socket, streaming):
+    """
+    Test a scenario where there is voice activity detected between silences.
+    :return:
+    """
+    speech_chunk_count = 5
+    probabilities = [0.0] * 5 + [1.0] * speech_chunk_count + [0.0] * 5
+    await simulate_streaming_with_probabilities(streaming, probabilities)
+
+    audio_out_socket.send.assert_called_once()
+    data = audio_out_socket.send.call_args[0][0]
+    assert isinstance(data, bytes)
+    # each sample has 512 frames of 4 bytes, expecting 7 chunks (5 with speech, 2 as padding)
+    assert len(data) == 512 * 4 * (speech_chunk_count + 2)
+
+
+@pytest.mark.asyncio
+async def test_voice_activity_short_pause(audio_in_socket, audio_out_socket, streaming):
+    """
+    Test a scenario where there is a short pause between speech, checking whether it ignores the
+    short pause.
+    """
+    speech_chunk_count = 5
+    probabilities = (
+        [0.0] * 5 + [1.0] * speech_chunk_count + [0.0] + [1.0] * speech_chunk_count + [0.0] * 5
+    )
+    await simulate_streaming_with_probabilities(streaming, probabilities)
+
+    audio_out_socket.send.assert_called_once()
+    data = audio_out_socket.send.call_args[0][0]
+    assert isinstance(data, bytes)
+    # Expecting 13 chunks (2*5 with speech, 1 pause between, 2 as padding)
+    assert len(data) == 512 * 4 * (speech_chunk_count * 2 + 1 + 2)
+
+
+@pytest.mark.asyncio
+async def test_no_data(audio_in_socket, audio_out_socket, streaming):
+    """
+    Test a scenario where there is no data received. This should not cause errors.
+    """
+    audio_in_poller = AsyncMock()
+    audio_in_poller.poll.return_value = None
+    streaming.audio_in_poller = audio_in_poller
+
+    await streaming.run()
+
+    audio_out_socket.send.assert_not_called()
+    assert len(streaming.audio_buffer) == 0
diff --git a/test/unit/agents/transcription/test_speech_recognizer.py b/test/unit/agents/transcription/test_speech_recognizer.py
new file mode 100644
index 0000000..88a5ac2
--- /dev/null
+++ b/test/unit/agents/transcription/test_speech_recognizer.py
@@ -0,0 +1,36 @@
+import numpy as np
+
+from control_backend.agents.transcription import SpeechRecognizer
+from control_backend.agents.transcription.speech_recognizer import OpenAIWhisperSpeechRecognizer
+
+
+def test_estimate_max_tokens():
+    """Inputting one minute of audio, assuming 300 words per minute, expecting 400 tokens."""
+    audio = np.empty(shape=(60 * 16_000), dtype=np.float32)
+
+    actual = SpeechRecognizer._estimate_max_tokens(audio)
+
+    assert actual == 400
+    assert isinstance(actual, int)
+
+
+def test_get_decode_options():
+    """Check whether the right decode options are given under different scenarios."""
+    audio = np.empty(shape=(60 * 16_000), dtype=np.float32)
+
+    # With the defaults, it should limit output length based on input size
+    recognizer = OpenAIWhisperSpeechRecognizer()
+    options = recognizer._get_decode_options(audio)
+
+    assert "sample_len" in options
+    assert isinstance(options["sample_len"], int)
+
+    # When explicitly enabled, it should limit output length based on input size
+    recognizer = OpenAIWhisperSpeechRecognizer(limit_output_length=True)
+    options = recognizer._get_decode_options(audio)
+
+    assert "sample_len" in options
+    assert isinstance(options["sample_len"], int)
+
+    # When disabled, it should not limit output length based on input size
+    assert "sample_rate" not in options
diff --git a/test/unit/conftest.py b/test/unit/conftest.py
index d7c10f2..ecf00c1 100644
--- a/test/unit/conftest.py
+++ b/test/unit/conftest.py
@@ -11,6 +11,7 @@ def pytest_configure(config):
     mock_spade = MagicMock()
     mock_spade.agent = MagicMock()
     mock_spade.behaviour = MagicMock()
+    mock_spade.message = MagicMock()
     mock_spade_bdi = MagicMock()
     mock_spade_bdi.bdi = MagicMock()
 
@@ -21,6 +22,7 @@ def pytest_configure(config):
     sys.modules["spade"] = mock_spade
     sys.modules["spade.agent"] = mock_spade.agent
     sys.modules["spade.behaviour"] = mock_spade.behaviour
+    sys.modules["spade.message"] = mock_spade.message
     sys.modules["spade_bdi"] = mock_spade_bdi
     sys.modules["spade_bdi.bdi"] = mock_spade_bdi.bdi
 
@@ -33,3 +35,26 @@ def pytest_configure(config):
     mock_config_module.settings = MagicMock()
 
     sys.modules["control_backend.core.config"] = mock_config_module
+
+    # --- Mock torch and zmq for VAD ---
+    mock_torch = MagicMock()
+    mock_zmq = MagicMock()
+    mock_zmq.asyncio = mock_zmq
+
+    # In individual tests, these can be imported and the return values changed
+    sys.modules["torch"] = mock_torch
+    sys.modules["zmq"] = mock_zmq
+    sys.modules["zmq.asyncio"] = mock_zmq.asyncio
+
+    # --- Mock whisper ---
+    mock_whisper = MagicMock()
+    mock_mlx = MagicMock()
+    mock_mlx.core = MagicMock()
+    mock_mlx_whisper = MagicMock()
+    mock_mlx_whisper.transcribe = MagicMock()
+
+    sys.modules["whisper"] = mock_whisper
+    sys.modules["mlx"] = mock_mlx
+    sys.modules["mlx.core"] = mock_mlx
+    sys.modules["mlx_whisper"] = mock_mlx_whisper
+    sys.modules["mlx_whisper.transcribe"] = mock_mlx_whisper.transcribe
diff --git a/uv.lock b/uv.lock
index 07ec3c1..c2bb61a 100644
--- a/uv.lock
+++ b/uv.lock
@@ -1332,6 +1332,7 @@ source = { virtual = "." }
 dependencies = [
     { name = "fastapi", extra = ["all"] },
     { name = "mlx-whisper", marker = "sys_platform == 'darwin'" },
+    { name = "numpy" },
     { name = "openai-whisper" },
     { name = "pyaudio" },
     { name = "pydantic" },
@@ -1358,6 +1359,7 @@ integration-test = [
     { name = "soundfile" },
 ]
 test = [
+    { name = "numpy" },
     { name = "pytest" },
     { name = "pytest-asyncio" },
     { name = "pytest-cov" },
@@ -1368,6 +1370,7 @@ test = [
 requires-dist = [
     { name = "fastapi", extras = ["all"], specifier = ">=0.115.6" },
     { name = "mlx-whisper", marker = "sys_platform == 'darwin'", specifier = ">=0.4.3" },
+    { name = "numpy", specifier = ">=2.3.3" },
     { name = "openai-whisper", specifier = ">=20250625" },
     { name = "pyaudio", specifier = ">=0.2.14" },
     { name = "pydantic", specifier = ">=2.12.0" },
@@ -1392,6 +1395,7 @@ dev = [
 ]
 integration-test = [{ name = "soundfile", specifier = ">=0.13.1" }]
 test = [
+    { name = "numpy", specifier = ">=2.3.3" },
     { name = "pytest", specifier = ">=8.4.2" },
     { name = "pytest-asyncio", specifier = ">=1.2.0" },
     { name = "pytest-cov", specifier = ">=7.0.0" },