Agent-to-Agent (A2A) Protocol & Federated Swarm Expert (2026 Edition)

English | Bahasa Indonesia

---

<a name="english"></a>

English

Orchestration & Integration

• Multi-Agent & MCP Ecosystem: multi-agent-orchestration, mcp-server-architect, mcp-sampling-critic, parallel-tool-orchestrator, durable-agent-execution-expert, computer-use-os-agent-expert
• Agentic Economy & Settlement: agentic-m2m-settlement, agentic-micro-economy-architect, doku-mcp-server
• Security & Contracts: structured-output-contracts, prompt-injection-firewall, zero-trust-secret-vault
• Lifecycle & Governance: brainstorming, zero-to-prod-orchestrator, session-memory-manager

Description

Production guide for implementing the open Agent-to-Agent (A2A) Protocol alongside Model Context Protocol (MCP). While MCP connects an agent to deterministic tools and resources (Agent-to-Tool), A2A enables opaque autonomous agents built on different frameworks (LangGraph, Pydantic AI, Google ADK, OpenAI Agents SDK, Mastra) to discover capabilities via /.well-known/agent.json Agent Cards, negotiate multi-turn tasks (tasks/send, tasks/sendSubscribe), exchange typed Artifacts, and settle M2M budgets.

Trigger Conditions

• Exposing an autonomous agent service to external partner agents via standard HTTP JSON-RPC + SSE.
• Publishing or consuming /.well-known/agent.json Agent Cards for dynamic capability discovery.
• Orchestrating cross-organization or polyglot multi-agent workflows where subagents do not share internal memory.
• Combining A2A horizontal agent delegation with MCP vertical tool execution.

---

A2A vs MCP Architectural Separation

Dimension	Model Context Protocol (MCP v1.x)	Agent-to-Agent Protocol (A2A)
Primary Relationship	Agent $\leftrightarrow$ Deterministic Tool / Resource	Client Agent $\leftrightarrow$ Remote Autonomous Agent
Discovery Mechanism	tools/list, resources/list, prompts/list	GET /.well-known/agent.json (Agent Card)
Execution Lifecycle	Synchronous RPC call (tools/call)	Stateful Task lifecycle (submitted $\rightarrow$ working $\rightarrow$ input-required $\rightarrow$ completed)
State & Reasoning	Stateless function execution	Opaque internal reasoning, memory, and multi-step planning

---

Core Implementation Patterns

1. Standard /.well-known/agent.json Agent Card Specification

// src/a2a/agent-card.ts
export interface A2AAgentSkill {
  id: string;
  name: string;
  description: string;
  tags: string[];
  inputModes: ('text/plain' | 'application/json')[];
  outputModes: ('text/plain' | 'application/json')[];
}

export interface A2AAgentCard {
  name: string;
  description: string;
  url: string;
  version: string;
  capabilities: {
    streaming: boolean;
    pushNotifications: boolean;
    stateTransitionHistory: boolean;
  };
  authentication: {
    schemes: ('Bearer' | 'OAuth2' | 'MutualTLS')[];
  };
  skills: A2AAgentSkill[];
}

export const FINTECH_AUDITOR_AGENT_CARD: A2AAgentCard = {
  name: 'vibes-fintech-auditor',
  description: 'Autonomous SNAP BI payment reconciliation and Z3 invariant verifier agent',
  url: 'https://agents.internal.example.com/a2a/fintech-auditor',
  version: '4.4.2',
  capabilities: {
    streaming: true,
    pushNotifications: true,
    stateTransitionHistory: true,
  },
  authentication: {
    schemes: ['Bearer'],
  },
  skills: [
    {
      id: 'verify-settlement-batch',
      name: 'Verify Settlement Batch',
      description: 'Audits DOKU SNAP BI webhook records against double-entry ledger postings',
      tags: ['fintech', 'doku', 'ledger', 'audit'],
      inputModes: ['application/json'],
      outputModes: ['application/json'],
    },
  ],
};

---

2. A2A JSON-RPC Task Handler (tasks/send & State Machine)

// src/a2a/task-router.ts
import { Hono } from 'hono';
import { z } from 'zod';
import { FINTECH_AUDITOR_AGENT_CARD } from './agent-card';

const app = new Hono();

const TaskSendSchema = z.object({
  jsonrpc: z.literal('2.0'),
  id: z.union([z.string(), z.number()]),
  method: z.enum(['tasks/send', 'tasks/get', 'tasks/cancel']),
  params: z.object({
    id: z.string().uuid(),
    sessionId: z.string().optional(),
    message: z.object({
      role: z.enum(['user', 'agent']),
      parts: z.array(
        z.object({
          type: z.enum(['text', 'data']),
          text: z.string().optional(),
          data: z.record(z.unknown()).optional(),
        })
      ),
    }),
  }),
});

app.get('/.well-known/agent.json', (c) => c.json(FINTECH_AUDITOR_AGENT_CARD));

app.post('/a2a/fintech-auditor', async (c) => {
  const body = await c.req.json();
  const parsed = TaskSendSchema.safeParse(body);
  if (!parsed.success) {
    return c.json({
      jsonrpc: '2.0',
      id: body?.id ?? null,
      error: { code: -32602, message: 'Invalid A2A task parameters', data: parsed.error.issues },
    }, 400);
  }

  const { id: rpcId, params } = parsed.data;

  return c.json({
    jsonrpc: '2.0',
    id: rpcId,
    result: {
      id: params.id,
      status: {
        state: 'completed',
        timestamp: new Date().toISOString(),
      },
      artifacts: [
        {
          name: 'audit-verdict.json',
          parts: [
            {
              type: 'data',
              data: { verified: true, discrepanciesFound: 0, taskId: params.id },
            },
          ],
        },
      ],
    },
  });
});

export default app;

---

Verification Checklist

• Agent service publishes a valid /.well-known/agent.json endpoint declaring authentication schemes and typed skills.
• All A2A task states adhere to standard transitions (submitted $\rightarrow$ working $\rightarrow$ input-required | completed | failed | canceled).
• Incoming parts payloads from remote agents are validated via Zod schemas and sanitized through prompt-injection-firewall.
• Long-running tasks support SSE streaming (tasks/sendSubscribe) or signed webhook push notifications.

---

<a name="bahasa-indonesia"></a>

Bahasa Indonesia

Integrasi Orkestrasi

• Ekosistem Multi-Agen & MCP: multi-agent-orchestration, mcp-server-architect, mcp-sampling-critic, parallel-tool-orchestrator, durable-agent-execution-expert, computer-use-os-agent-expert
• Ekonomi & Settlement Agen: agentic-m2m-settlement, agentic-micro-economy-architect, doku-mcp-server
• Keamanan & Kontrak: structured-output-contracts, prompt-injection-firewall, zero-trust-secret-vault
• Siklus Hidup & Tata Kelola: brainstorming, zero-to-prod-orchestrator, session-memory-manager

Deskripsi

Panduan produksi untuk mengimplementasikan Agent-to-Agent (A2A) Protocol berdampingan dengan Model Context Protocol (MCP). Jika MCP menghubungkan agen ke tool deterministik (Agent-to-Tool), A2A memungkinkan agen otonom lintas framework untuk saling menemukan kapabilitas melalui /.well-known/agent.json (Agent Card), menegosiasikan tugas multi-giliran (tasks/send), dan bertukar Artifact terstruktur.

Panduan Arsitektur Inti

1. Sinergi A2A + MCP: Gunakan A2A untuk delegasi horizontal antar-agen spesialis, dan gunakan MCP untuk pemanggilan fungsi/tool vertikal di dalam masing-masing agen.
2. Penemuan Dinamis (Agent Card): Publikasikan kapabilitas, mode input/output, dan skema autentikasi di /.well-known/agent.json.
3. Pertahanan Injeksi Lintas-Agen: Perlakukan setiap pesan dari agen eksternal sebagai input yang tidak tepercaya; validasi dengan skema JSON ketat dan prompt-injection-firewall.

Daftar Periksa Verifikasi (Checklist)

• Endpoint /.well-known/agent.json tersedia dan mendeklarasikan skema autentikasi.
• Siklus hidup tugas mengikuti status standar (submitted, working, input-required, completed, failed).
• Payload antar-agen divalidasi dengan Zod dan disanitasi dari serangan indirect prompt injection.