2nd/10_Wiki/Topics/Architecture/A2A.md

---
id: wiki-2026-0508-a2a
title: A2A (Agent-to-Agent Protocol)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Agent-to-Agent, A2A Protocol, Agent2Agent]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [agents, protocol, interop, anthropic, mcp]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: anthropic-sdk
---

# A2A (Agent-to-Agent Protocol)

## 매 한 줄
> **"매 agent 가 다른 agent 와 모델/벤더 차이 없이 task 를 위임/협상/결과 교환 하기 위한 open protocol."** Google 이 2025 봄 announce, 2025-06 Linux Foundation 으로 stewardship 이전, 2026 현재 Anthropic / Microsoft / Salesforce 등 50+ 기업 adoption. 매 MCP (tool/data) 와 directly complementary — 매 A2A 는 *agent ↔ agent*, MCP 는 *agent ↔ tool*.

## 매 핵심

### 매 5 design principles
- **Agentic by default**: 매 agent autonomy 를 가정 (매 mere RPC 가 아님).
- **Modality-agnostic**: 매 text / audio / video / structured data 모두 transport.
- **Built on web standards**: HTTP + JSON-RPC 2.0 + SSE / WebSocket — 매 separate runtime 불필요.
- **Secure by design**: OAuth 2.1, 매 mTLS, 매 capability scoping.
- **Long-running task aware**: 매 minutes ~ days 의 async task — 매 polling + push-notification 모두 지원.

### 매 핵심 primitives
- **AgentCard** (`/.well-known/agent.json`): 매 agent 의 capability advertisement.
- **Task**: 매 unit of work — `submitted → working → input-required → completed/failed/canceled`.
- **Message + Artifact**: 매 conversation chunk + 매 final output.
- **Streaming**: SSE 로 매 partial token / 매 status update 전송.

### 매 응용
1. Cross-vendor agent orchestration (Claude → Gemini → in-house).
2. Specialist agent dispatch (legal, finance, code-review).
3. Marketplace 의 agent invocation.

## 💻 패턴

### 1) AgentCard 발행
```json
{
  "name": "claude-research-agent",
  "version": "1.2.0",
  "url": "https://api.example.com/a2a",
  "capabilities": {
    "streaming": true,
    "pushNotifications": true,
    "stateTransitionHistory": true
  },
  "skills": [
    {"id": "deep-research", "description": "Multi-source web research", "inputModes": ["text"], "outputModes": ["text", "file"]}
  ],
  "auth": {"type": "oauth2.1", "scopes": ["task:submit"]}
}
```

### 2) Task 제출 (client agent)
```python
import httpx, uuid

resp = httpx.post(
    "https://api.example.com/a2a/tasks/send",
    json={
        "jsonrpc": "2.0", "id": "1", "method": "tasks/send",
        "params": {
            "id": str(uuid.uuid4()),
            "message": {
                "role": "user",
                "parts": [{"type": "text", "text": "Summarize Q4 earnings of NVDA."}]
            }
        }
    },
    headers={"Authorization": f"Bearer {token}"}
)
task = resp.json()["result"]
```

### 3) SSE streaming 으로 partial result 수신
```python
with httpx.stream("POST", url + "/tasks/sendSubscribe", json=req) as s:
    for line in s.iter_lines():
        if line.startswith("data:"):
            event = json.loads(line[5:])
            if event["type"] == "status": ...
            elif event["type"] == "artifact": print(event["artifact"]["parts"])
```

### 4) Push notification 등록 (long task)
```python
httpx.post(url + "/tasks/pushNotification/set", json={
    "taskId": task["id"],
    "pushNotificationConfig": {
        "url": "https://my-app.com/a2a/webhook",
        "token": webhook_secret
    }
})
```

### 5) Server-side handler (FastAPI 예)
```python
from fastapi import FastAPI
app = FastAPI()

@app.post("/a2a/tasks/send")
async def send(req: dict):
    task_id = req["params"]["id"]
    # 매 background worker 에게 dispatch
    await queue.put((task_id, req["params"]["message"]))
    return {"jsonrpc":"2.0","id":req["id"],"result":{"id":task_id,"status":{"state":"submitted"}}}
```

### 6) Multi-agent orchestration (A2A + MCP combo)
```python
# Orchestrator agent: A2A 로 specialist 호출, MCP 로 tool 사용
research = await a2a_call("research-agent", query)
draft = await claude.messages.create(  # MCP tools attached
    model="claude-opus-4-7",
    tools=mcp_tools,
    messages=[{"role":"user","content": f"Draft based on: {research}"}]
)
```

### 7) Capability negotiation
```python
card = httpx.get(agent_url + "/.well-known/agent.json").json()
if not card["capabilities"]["streaming"]:
    # 매 polling fallback
    use_polling = True
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 매 agent ↔ tool / data | MCP |
| 매 agent ↔ agent (cross-vendor) | A2A |
| 매 same-process agent | Direct call (no protocol) |
| 매 long-running (>30s) | A2A + push notification |
| 매 strict typing 필요 | A2A + JSON Schema in skill spec |

**기본값**: 매 cross-org / cross-vendor agent collaboration 에 매 A2A, 매 internal tool wiring 은 매 MCP.

## 🔗 Graph
- Adjacent: [[OAuth 2.1]]

## 🤖 LLM 활용
**언제**: 매 multi-vendor agent stack 의 interop, 매 long-running specialist agent 의 호출.
**언제 X**: 매 single-process tool 호출 (매 MCP 를 사용), 매 latency-critical (<50ms) 의 inner loop.

## ❌ 안티패턴
- **A2A 로 tool 호출**: 매 MCP scope. 매 protocol 혼동.
- **No AgentCard**: 매 capability 알려주지 않으면 매 client 가 fallback 못함.
- **Sync polling on long task**: 매 push notification 또는 SSE 필수.
- **Token leakage**: 매 webhook URL 의 token 검증 누락.

## 🧪 검증 / 중복
- Verified (a2aproject.org spec v0.3, Linux Foundation A2A Project 2025-06-23, Anthropic A2A blog 2025).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — A2A protocol primitives + MCP comparison + 7 working patterns |