Protocols & Standards
The protocol landscape for agent-to-service and agent-to-agent communication
Summary
Three protocols dominate: MCP (agent-to-tools/data, vertical), A2A (agent-to-agent, horizontal), ACP (agent commerce). Most production systems layer them: agents use MCP to access their tools, A2A to delegate to peers, and optionally ACP for purchases. Supporting standards (OpenTelemetry, llms.txt) add observability and API discovery.
Orchestrator (uses A2A)
├── MCP → tools
├── MCP → data
├── A2A → Specialist 1
│ └── MCP → Specialist's tools
└── A2A → Specialist 2
└── MCP → Specialist's tools- MCP 2025-11-25: tools, resources, prompts, async tasks
- A2A v1.0 RC: task delegation across organizational boundaries
- ACP beta: agent-initiated commerce with payment delegation
- Publish: /.well-known/agent.json, /.well-known/acp.json, /AGENTS.md, /llms.txt
- Transport: stdio (local), Streamable HTTP (remote)
Three mature protocols and several complementary standards define how agents discover capabilities, execute work, and integrate across organizational boundaries.
Quick Start
Does your agent need tools? → Use MCP (spec 2025-11-25). This is the baseline.
Multiple agents delegating work? → Add A2A (v1.0 RC) for peer-to-peer task delegation.
Agent initiating purchases? → Add ACP (beta) for commerce with payment boundaries.
Complex multi-step workflows? → Use Arazzo to choreograph API sequences.
Need observability? → Instrument with OpenTelemetry GenAI semconv — stable and widely adopted.
The Protocols
| Protocol | Purpose | Learn More |
|---|---|---|
| MCP 2025-11-25 | Agent-to-tool / agent-to-data (vertical) | MCP docs |
| A2A v1.0 RC | Agent-to-agent delegation (horizontal) | A2A docs |
| ACP beta | Agent-initiated commerce | ACP docs |
| Arazzo 1.0.1 | Multi-step API workflow sequences | Comparison page |
| OpenTelemetry GenAI | Observability & tracing | Emerging standards |
| AGENTS.md | Agent context documentation | Emerging standards |
| llms.txt | LLM-friendly API docs | Emerging standards |
Architecture Pattern
Most production systems use MCP + A2A + supporting standards:
Orchestrator Agent (uses A2A)
├── MCP → Internal tools
├── MCP → Data sources
├── A2A → Specialist Agent 1
│ └── MCP → Specialist's tools
└── A2A → Specialist Agent 2
└── MCP → Specialist's toolsDiscovery Checklist
Publish these endpoints for agents to find you:
/.well-known/agent.json— A2A AgentCard (for multi-agent scenarios)/.well-known/acp.json— ACP capabilities (if you accept agent-initiated payments)/AGENTS.md— Agent context & MCP servers (always useful)/llms.txtor/llms-full.txt— LLM-friendly docs- Structured data as JSON-LD (schema.org markup) for auto-discovery
Full Comparison & Decision Guide
See Protocol Comparison for maturity, transport details, auth models, and when to use each.
See Emerging Standards for Arazzo workflows, observability, semantic metadata, and draft specs to monitor.