Agent2Agent: agent cards and opaque collaboration
A2A lets independent agents discover each other and exchange work without exposing their internals. What an agent card declares, and what A2A is not.
The Agent2Agent (A2A) protocol is an open standard for communication between independent, potentially opaque AI agent systems, built on different frameworks by different vendors. It was originally developed by Google and donated to the Linux Foundation; the project page names a technical steering committee drawn from several companies, and the protocol is licensed under Apache 2.0. The specification read for this episode is version 1.0.
An agent card describes an agent: its skills, capabilities, the interfaces it serves, and the security schemes a client must use to authenticate. An agent may also offer an extended card with more detail to authenticated clients.
The project is explicit about what A2A is not: not an agent development kit, not a protocol for an agent's own sub-agents or tool calls (it points to MCP for tools), and not a replacement for MCP. Agents collaborate without sharing internal memory, plans or tools.
In a finance flow, a corporate's procurement agent could find a supplier's sales agent through its card, authenticate with the scheme it declares, and exchange quotes as structured messages.
Limits. An agent card is a self-description. It says how to authenticate, not whether the operator behind the agent is who it claims to be or is allowed to act for a customer. Opaque by design also means you cannot inspect the other side: trust has to come from authentication, mandates and records.
Sources: Agent2Agent (A2A) Protocol specification, A2A Protocol: home, governance and licence