← Back to all projects

A2A
The open protocol for one independent agent to discover and delegate work to another
OfficialApache-2.0
What is A2A?
A2A defines how separately built agents advertise their capabilities, exchange messages and track work that may outlive one connection. Version 1.0 separates a shared data model and operations from JSON-RPC, gRPC and HTTP/REST bindings, while official SDKs live in other repositories. Use it between autonomous services that need task state, streaming or asynchronous updates; a database query or other bounded tool call still belongs behind MCP, and adopting the specification does not supply an agent, a registry or authorization policy.
What can you do with A2A?
- Publish capabilities in an Agent Card — An A2A server exposes a JSON document describing its identity, service URL, supported input and output media, optional protocol extensions, security schemes and named skills. A client can decide whether the remote agent is suitable before sending it work, without learning its prompts, memory or internal tools.
- Keep long-running work in a Task — A request can return a stateful Task rather than a final message. The client can retrieve, list or cancel it, group related work with a context identifier, and receive the resulting documents or structured data as Artifacts while the remote agent keeps its implementation private.
- Choose immediate, streamed or disconnected updates — Simple interactions return a Message directly. Longer work can stream status and Artifact chunks over the selected protocol binding, or notify a webhook after the original client has disconnected, so human approval and external processing do not have to fit inside one HTTP response.
- Share semantics across three protocol bindings — Version 1.0 defines the canonical data model in Protocol Buffers, describes operations independently, then maps them to JSON-RPC, gRPC and HTTP/REST. Implementations can choose a transport without giving Task, Message, Agent Card and Artifact different meanings.
- Reuse web authentication instead of carrying credentials in messages — The Agent Card declares OAuth, OpenID Connect, API-key or mutual-TLS requirements, while credentials travel through the binding's normal headers and channels rather than inside A2A Messages. The protocol describes the scheme; each server still has to enforce who may discover an agent, read a Task or receive an Artifact.
- Put A2A above MCP, not in place of it — A remote agent may plan, keep state and use several tools while accepting delegated work through A2A. That agent can use MCP internally to reach databases, APIs and other bounded capabilities, so the two protocols occupy different links in the same system.
Before you choose A2A
- Agent Cards make discovery possible but do not provide a universal registry; well-known URLs, curated catalogues and private registries remain deployment choices.
- The specification names security schemes but does not implement access control, tenant isolation or credential storage; every server and client must supply those controls.
- The specification and official SDKs ship from separate repositories, so support for a new protocol version must be checked in the language SDK you plan to deploy.
Star history
17 Aug to 30 Oct · +973
25.4k26.4k
Frequently asked questions
Is A2A free for commercial use?
A2A is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.