Agent-to-Agent protocol v1.0 RC
What an agent-card.json buys you, what it doesn't, and how it relates to MCP and WebMCP.
A2A (Agent-to-Agent) sits on top of MCP. Where MCP is "model calls a tool", A2A is "agent calls another agent". Different layer, similar plumbing.
The signature artifact is agent-card.json. It lives at /.well-known/agent-card.json and describes what an agent can do as a list of skills. Each skill has an id, a description, examples, input modes, output modes. The other agent reads the card, picks a skill, sends a request.
That sounds like MCP again. The difference is intent. MCP tools are model-callable verbs on a domain. A2A skills are agent-callable capabilities at a higher level. A skill might fan out to multiple MCP tool calls. A skill carries semantic intent ("plan a trip", "review this code"), not just an action.
I ship agent-card.json on every domain I deploy because it costs nothing and shows up in agent registries. I have not yet seen a real-world A2A negotiation in production traffic. It is a forward bet. The spec is RC, the ecosystem is shallow, but the cost of supporting it is one route handler.
If you build today: ship MCP first (people use it now), ship A2A as a hygiene layer (so when a registry indexes your domain it sees the card), don't bet your roadmap on A2A becoming load-bearing for another two years.