​

Agent-to-Agent (A2A) Communication Guide

⚠️ OUTDATED (2026-06-20). This guide describes the retired Google ADK / InMemoryTaskManager / AgentCommunicationService model. The current design runs A2A over Temporal tasks + the workflow event stream. See:

• docs/adr/2026-06-20-a2a-transport-correctness.md — how an A2A task produces and returns a result (authoritative).
• docs/plans/2026-03-10-agent-delegation-remote-mcp.md — delegation via A2AAgentTool.
• docs/plans/2026-03-10-a2a-spec-compliance.md — A2A v0.3.0 wire contract.

The sections below are kept for historical context only.

• AgentCommunicationService – orchestration layer that issues tasks and tracks their results
• InMemoryTaskManager – default store used to register the delegated tasks
• Google ADK “ask_agent” tool – automatically injected into your LlmAgent so it can call helper agents from within its reasoning chain.

​

1. When to Use

​

2. Architecture Overview

Primary Agent ──ask_agent()──► AgentCommunicationService ──► TaskManager
                                               │
                                               ▼
                                EventBroker (TaskCreated / …)
                                               │
                                 Helper Agent Runner picks task
                                               │
                               (work, emit TaskCompleted / Failed)
                                               ▼
                              AgentCommunicationService resolves

• ask_agent tool → validates input, generates a unique task-id, calls TaskManager (tasks/send) with metadata:
  {
    "agent_id": "<helper-agent-uuid>",
    "is_agent_to_agent": true
  }
  
• Service stores an asyncio.Event and waits (max configurable 60 s) for TaskCompleted/Failed.
• Result (or error) is transparently returned to the calling agent so it can continue its reasoning.

​

3. Prerequisites

• AgentArea backend running (uvicorn core.agentarea.main:app --reload)
• Google ADK installed (pip install google-adk)
• Event broker configured (Redis/Kafka) – required for task status events
• Agents you want to communicate must exist in the database (see CLI or REST /agents/).

​

4. Enabling / Disabling Communication

task_execution_service.execute_task_async(
    ...,
    metadata={
        "enable_agent_communication": false   # disable for this task
    }
)

​

5. Using the ask_agent Tool (inside an agent)

​

Tool specification (autogenerated)

{
  "task_id": "b4f5…",
  "status": "completed",
  "response": "… helper agent reply …"
}

​

Example Prompt

If you cannot solve the maths problem yourself, call ask_agent with:
{
  "agent_id": "77fa25d0-bd50-466c-8599-54f6f61c6d1b",
  "message": "Compute the integral of sin(x)/x from 0 to infinity."
}

​

6. Manual Task Creation (REST)

curl -X POST http://localhost:8000/v1/tasks/ \
  -H "Content-Type: application/json" \
  -d '{
        "message": "Translate this to French: Hello World",
        "agent_id": "77fa25d0-bd50-466c-8599-54f6f61c6d1b",
        "metadata": { "is_agent_to_agent": true }
      }'

GET /v1/tasks/{task_id}

​

7. Full Example Script

1. Creating primary & helper agents
2. Primary agent delegating a math task
3. Waiting for completion
4. Local simulation without backend (Google ADK only)

python examples/agent_to_agent_example.py

​

8. Configuration Options

​

9. Troubleshooting

export LOGLEVEL=debug

​

10. Security Considerations

• All delegated tasks inherit user_id of the caller (if provided) – use RBAC to restrict cross-agent calls.
• Validate metadata size to avoid abuse.
• When exposing externally, protect the /tasks/send endpoint with authentication.