Summary

Two protocols dominate: (agent-to-tools/data, vertical) and (agent-to-agent, horizontal). Most production systems layer them: agents use MCP to access their tools and A2A to delegate to peers. Commerce protocols (, ) that ride on top of these for agent-initiated purchases now have their own section - see Agentic Commerce. Supporting standards (OpenTelemetry, ) add observability and discovery.

Orchestrator (uses A2A)
  ├── MCP → tools
  ├── MCP → data
  ├── A2A → Specialist 1
  │    └── MCP → Specialist's tools
  └── A2A → Specialist 2
       └── MCP → Specialist's tools

MCP: remote tools, resources, and prompts - 2026-07-28 is the current revision and is supported by all Tier 1 ; async Tasks are an official extension
: page-scoped browser tools through document.modelContext; a W3C Community Group draft, not remote MCP
A2A v1.0.1: production-grade task delegation across organizational boundaries
Commerce (UCP, ACP): agent-initiated purchases - see Agentic Commerce, not covered on this page
Publish: /.well-known/agent-card.json, /AGENTS.md, /llms.txt
Transport: stdio (local), Streamable HTTP (remote)

These protocols and complementary standards define how agents discover capabilities, execute work, and integrate across organizational boundaries.

Quick Start

Does your agent need remote tools or data? → Use MCP. 2026-07-28 is the current revision, with a stateless core and no sessions, and all Tier 1 SDKs support it. Retain 2025-11-25 compatibility only when a real client still requires it.

Does a page need to expose tools while an agent visits it? → Use WebMCP. It is a browser proposal built around document.modelContext; it does not require or replace the remote MCP wire protocol.

Multiple agents delegating work? → Add A2A (v1.0.1) for peer-to-peer task delegation.

Agent initiating purchases? → See Agentic Commerce for UCP and ACP, payment authority, and trust and risk. Note that "ACP" is overloaded with two unrelated protocols; the Checkout Protocols page disambiguates them.

Complex multi-step workflows? → Use Arazzo to choreograph API sequences.

Need observability? → Instrument with OpenTelemetry GenAI semconv - widely adopted, but still marked Development.

The Protocols

Protocol	Purpose	Learn More
MCP 2026-07-28	Agent-to-tool / agent-to-data (vertical)	MCP docs
WebMCP draft	Page-scoped browser tools	WebMCP docs
A2A v1.0.1	Agent-to-agent delegation (horizontal)	A2A docs
UCP / ACP (commerce)	Agent-initiated commerce	Agentic Commerce
1.1.0	Multi-step API workflow sequences	Comparison page
OpenTelemetry GenAI	Observability &	Testing: Observability
	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 tools

Discovery Checklist

Publish these endpoints for agents to find you:

/.well-known/agent-card.json - A2A (for multi-agent scenarios)
/.well-known/ucp or /.well-known/acp.json - commerce capabilities (if you accept agent-initiated payments) - see Agentic Commerce
/AGENTS.md - Agent context & MCP servers (always useful)
/llms.txt or /llms-full.txt - LLM-friendly docs
Structured data as (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 draft specs to monitor.

Paid resources and structured discovery

Payments: , , mandates, and spending controls for paid resources and agent-initiated purchases.
Agent Catalogs and Structured Feeds: , , and feed/query boundaries.

Protocols & Standards