The system
How the system worksThe DirectoryDispatchThe LedgerAI modelsAI agentsYour installationThe appThe research libraryWhat changes for your businessUse it from your softwareEverything it can do
Messaging
iMessageWhatsApp
Ads and leads
AdsCreativeLeads
Devices
Devices
Industries
RealtorsRestaurantsHome services
Explore
How it worksPermissions and limitsCompareSet upQuestionsText me

A2A, from its own documents

Agent2Agent is the protocol for the layer above tools: agents talking to other agents. Its repository states the problem it exists to solve — enabling "gen AI agents built on diverse frameworks by different companies running on separate

A2A, from its own documents

Agent2Agent is the protocol for the layer above tools: agents talking to other agents. Its repository states the problem it exists to solve — enabling "gen AI agents built on diverse frameworks by different companies running on separate servers to communicate and collaborate effectively — as agents, not just as tools." Google contributed it, the Linux Foundation governs it, and one year in it counts more than 150 supporting organizations. This page is the reference record, from the protocol's own repository and its foundation's announcements.

The one-line definition

The repo describes itself as "an open protocol enabling communication and interoperability between opaque agentic applications." The word doing the work is opaque. A2A's core design commitment is that agents cooperate without exposing internal state, memory, or tools — you interact with what an agent claims it can do, never with how it does it.

Mechanically it is deliberately boring: JSON-RPC 2.0 over HTTPS, with synchronous request/response, streaming over server-sent events, and asynchronous push notifications. Apache 2.0 licensed, 25,000 GitHub stars. Boring is the point; protocols win by being unobjectionable.

The agent card is the whole discovery model

An A2A agent publishes an agent card: a structured document describing its capabilities and connection details. Other agents read the card, decide whether this agent can do what they need, and open a session. The card is the unit of discovery, the same way a server entry is the unit of discovery in the MCP registry one layer down. Capability description moves to a machine-readable public record; negotiation happens against the record, not the implementation.

One year to 150 organizations

The Linux Foundation's April 9, 2026 announcement marked the protocol's first year: "more than 150 organizations supporting the standard," with production deployments on Google Cloud, Microsoft Azure (AI Foundry and Copilot Studio), and AWS (Bedrock AgentCore Runtime).

All three hyperscalers shipping the same agent-coordination protocol within a year is the notable fact. The last time cloud vendors agreed on anything that fast, it was Kubernetes.

The division of labor with MCP

The foundation states the boundary in one sentence: "A2A defines how agents communicate and coordinate with each other across organizational boundaries, while MCP defines how agents connect to internal tools and data sources." Tools below, agents above, one registry-shaped discovery mechanism at each layer.

What neither layer covers, and A2A's enterprise-readiness language only gestures at: how much authority travels with a delegated task. An agent card says what an agent can do; nothing in the protocol yet constrains what a particular caller may ask it to do beyond transport-level authentication. That open edge — delegation with enforceable limits — is where the current research literature points, and it is the axis on which agent-to-agent systems will differentiate next.

References (2)
  1. a2aproject/A2A (GitHub)“An open protocol enabling communication and interoperability between opaque agentic applications.”
  2. A2A Protocol Surpasses 150 Organizations (Linux Foundation, Apr 9 2026)“A2A defines how agents communicate and coordinate with each other across organizational boundaries, while MCP defines how agents connect to internal tools and data sources.”