What is A2A Protocol?

Updated on

A2A Protocol (Agent-to-Agent Protocol) is an open standard for discovery, messaging, task lifecycle management, and artifact exchange between independent, potentially opaque agent systems.

Quick Facts

Full NameAgent-to-Agent Protocol
Created2025 by Google
SpecificationOfficial Specification

How It Works

Data model, operations, and bindings

The canonical model includes AgentCard, Message, Part, Task, Artifact, Extension, and related security metadata. Abstract operations cover sending a message, streaming, retrieving, listing, canceling, and subscribing to tasks, plus push-notification configuration. Bindings map those semantics to JSON-RPC, gRPC, HTTP+JSON, or future mechanisms. Applications should pin a major protocol version and regenerate derived SDK or schema artifacts from the normative definition.

Discovery is not trust

An Agent Card advertises identity metadata, skills, media modes, interfaces, protocol versions, and security schemes. Clients must fetch it from an expected origin, validate every selected interface, and apply an organizational trust policy. Signed Agent Cards can provide integrity and signer evidence only when the verification key and provenance are trusted. A skill description is a capability claim, not authorization, uptime, quality, or compliance proof.

Messages, tasks, and asynchronous results

Send Message may return a direct Message for a simple exchange or a stateful Task for longer work. A Task carries identity, status, history, and Artifacts and can progress through terminal and nonterminal states. Clients may poll, stream status and artifact events, or receive authenticated push notifications according to the selected binding. They must deduplicate events, validate terminal state, bound history, and reject updates for an inaccessible task or tenant.

Security and application responsibility

Transport authentication identifies a caller or peer; the server still authorizes the requested operation, tenant, data, and side effect. Propagate user authority deliberately, minimize delegated context, and do not forward reusable credentials. Use deadlines, idempotency keys, replay defenses, audit IDs, and reconciliation for writes. Treat remote messages and artifacts as untrusted input, and evaluate semantic interoperability, failure recovery, and policy enforcement in addition to wire conformance.

Key Characteristics

  • Defines a canonical model for Agent Cards, messages, tasks, parts, and artifacts
  • Separates binding-independent operations from JSON-RPC, gRPC, and HTTP+JSON mappings
  • Supports direct responses, stateful tasks, streaming, polling, and push notifications
  • Allows agents to remain opaque while advertising bounded skills and interfaces
  • Provides version and security metadata but leaves trust policy to implementations
  • Complements MCP without replacing tool access, authorization, or workflow control

Common Use Cases

  1. Delegating a long-running task to an agent owned by another service or organization
  2. Connecting specialized agents implemented with different languages or frameworks
  3. Exchanging typed messages and artifacts while preserving private internal execution
  4. Tracking asynchronous agent work through polling, streams, or push notifications
  5. Building gateways that enforce tenant, identity, version, and policy boundaries

Example

loading...
Loading code...

Frequently Asked Questions

How is A2A different from MCP?

A2A standardizes communication and task delegation between independent agents. MCP standardizes how an application exposes tools, resources, and prompts to an AI host or client. A remote agent may use MCP internally and communicate over A2A externally. Neither protocol decides whether a business action is authorized or correct.

Is A2A still a Google protocol?

Google originally developed and launched A2A in 2025, then donated the project to the Linux Foundation. The current protocol is governed as an open, vendor-neutral project with a technical steering committee. Historical origin and current governance are separate facts, so implementations should follow the versioned A2A specification rather than an early launch post.

Does A2A use only HTTP, JSON-RPC, and SSE?

No. Early versions were commonly described that way. A2A 1.0 separates its canonical data model and abstract operations from bindings, with core mappings for JSON-RPC, gRPC, and HTTP+JSON. Streaming and asynchronous delivery depend on the chosen binding and capabilities. Pin the specification and negotiate a compatible interface.

Does an Agent Card prove that an agent is trusted?

No. An Agent Card advertises claims about identity, skills, interfaces, versions, and security. Fetch it from an expected origin, validate its fields, and apply trust policy. A valid signature proves integrity and signer possession only relative to a trusted verification key; it does not grant authorization or prove output quality.

Does an A2A Task guarantee exactly-once side effects?

No. Task identity and lifecycle improve tracking, but retries, disconnects, and partial failures can leave an external write with an unknown outcome. Use application-level idempotency keys, authoritative status checks, deduplicated events, and reconciliation. A terminal protocol state should reflect verified business state rather than replace it.

Related Terms

Related Articles