diff --git a/README.md b/README.md
index 4c9a3008..8b2e485c 100644
--- a/README.md
+++ b/README.md
@@ -301,6 +301,51 @@ Authorization: Bearer your-token-jwt
 - **LangGraph**: Framework for building stateful, multi-agent workflows
 - **ReactFlow**: Library for building node-based visual workflows
 
+## 📊 Langfuse Integration (Tracing & Observability)
+
+Evo AI platform natively supports integration with [Langfuse](https://langfuse.com/) for detailed tracing of agent executions, prompts, model responses, and tool calls, using the OpenTelemetry (OTel) standard.
+
+### Why use Langfuse?
+
+- Visual dashboard for agent traces, prompts, and executions
+- Detailed analytics for debugging and evaluating LLM apps
+- Easy integration with Google ADK and other frameworks
+
+### How it works
+
+- Every agent execution (including streaming) is automatically traced via OpenTelemetry spans
+- Data is sent to Langfuse, where it can be visualized and analyzed
+
+### How to configure
+
+1. **Set environment variables in your `.env`:**
+
+   ```env
+   LANGFUSE_PUBLIC_KEY="pk-lf-..."   # Your Langfuse public key
+   LANGFUSE_SECRET_KEY="sk-lf-..."   # Your Langfuse secret key
+   OTEL_EXPORTER_OTLP_ENDPOINT="https://cloud.langfuse.com/api/public/otel" # (or us.cloud... for US region)
+   ```
+
+   > **Attention:** Do not swap the keys! `pk-...` is public, `sk-...` is secret.
+
+2. **Automatic initialization**
+
+   - Tracing is automatically initialized when the application starts (`src/main.py`).
+   - Agent execution functions are already instrumented with spans (`src/services/agent_runner.py`).
+
+3. **View in the Langfuse dashboard**
+   - Access your Langfuse dashboard to see real-time traces.
+
+### Troubleshooting
+
+- **401 Error (Invalid credentials):**
+  - Check if the keys are correct and not swapped in your `.env`.
+  - Make sure the endpoint matches your region (EU or US).
+- **Context error in async generator:**
+  - The code is already adjusted to avoid OpenTelemetry context issues in async generators.
+- **Questions about integration:**
+  - See the [official Langfuse documentation - Google ADK](https://langfuse.com/docs/integrations/google-adk)
+
 ## 🤖 Agent 2 Agent (A2A) Protocol Support
 
 Evo AI implements the Google's Agent 2 Agent (A2A) protocol, enabling seamless communication and interoperability between AI agents. This implementation includes:
diff --git a/src/services/a2a_task_manager.py b/src/services/a2a_task_manager.py
index 6f74aef7..617171e6 100644
--- a/src/services/a2a_task_manager.py
+++ b/src/services/a2a_task_manager.py
@@ -3,6 +3,7 @@ import asyncio
 from collections.abc import AsyncIterable
 from typing import Dict, Optional
 from uuid import UUID
+import json
 
 from sqlalchemy.orm import Session
 
@@ -306,7 +307,7 @@ class A2ATaskManager:
             external_id = task_params.sessionId
             full_response = ""
 
-            # We use the same streaming function used in the WebSocket
+            # Use the same pattern as chat_routes.py: deserialize each chunk
             async for chunk in run_agent_stream(
                 agent_id=str(agent.id),
                 external_id=external_id,
@@ -316,9 +317,14 @@ class A2ATaskManager:
                 memory_service=memory_service,
                 db=self.db,
             ):
-                # Send incremental progress updates
-                update_text_part = {"type": "text", "text": chunk}
-                update_message = Message(role="agent", parts=[update_text_part])
+                try:
+                    chunk_data = json.loads(chunk)
+                except Exception as e:
+                    logger.warning(f"Invalid chunk received: {chunk} - {e}")
+                    continue
+
+                # The chunk_data must be a dict with 'type' and 'text' (or other expected format)
+                update_message = Message(role="agent", parts=[chunk_data])
 
                 # Update the task with each intermediate message
                 await self.update_store(
@@ -337,24 +343,24 @@ class A2ATaskManager:
                         final=False,
                     ),
                 )
-                full_response += chunk
+                # If it's text, accumulate for the final response
+                if chunk_data.get("type") == "text":
+                    full_response += chunk_data.get("text", "")
 
-            # Determine the task state
+            # Determine the final state of the task
             task_state = (
                 TaskState.INPUT_REQUIRED
                 if "MISSING_INFO:" in full_response
                 else TaskState.COMPLETED
             )
 
-            # Create the final response part
+            # Create the final response
             final_text_part = {"type": "text", "text": full_response}
             parts = [final_text_part]
             final_message = Message(role="agent", parts=parts)
-
-            # Create the final artifact from the final response
             final_artifact = Artifact(parts=parts, index=0)
 
-            # Update the task in the store with the final response
+            # Update the task with the final response
             await self.update_store(
                 task_params.id,
                 TaskStatus(state=task_state, message=final_message),