What is the A2A (Agent2Agent) protocol?

Updated 3 November 2026 · Editorial assessment

What does A2A standardize?

A2A establishes a common interface between a client and an agent server. The client may itself be another agent. A server publishes its capabilities and accepts work through a supported protocol binding. This lets teams connect separately developed agent systems without making their internal implementation identical. The boundary is the contract they expose, not whether they use the same model, programming language or framework.

Its central concepts are Agent Cards, messages, tasks and artifacts. A message carries an interaction; a task represents work with a lifecycle; an artifact represents an output. A context identifier groups related interactions while a task identifier distinguishes a particular unit of work. These identifiers help applications correlate activity, but they do not create a shared memory store or decide how long an enterprise must retain records.

Source: a2a-protocol.org/latest/specification · Updated 3 November 2026

How do Agent Cards support discovery?

An Agent Card describes an agent, its skills, supported interfaces, input and output types, capabilities and security requirements. The standard well-known location is /.well-known/agent-card.json. A caller can also obtain a card through an approved catalog or a known URL. The specification does not require a universal public registry, and discovering a card is not permission to invoke the endpoint.

For procurement, ask to see the actual card returned by the production endpoint. Compare its advertised skills with the task you intend to delegate. Check supported interface versions and transports, rather than assuming a product label proves compatibility. The card's agent version and an interface's protocol version are different fields. An updated agent can still expose an older interface, which matters when choosing client libraries.

Source: a2a-protocol.org/latest/specification · Updated 3 November 2026

Comparison: What is the A2A (Agent2Agent) protocol?
Contract elementPurposeEnterprise check
Agent CardDiscover skills and interfacesApproved endpoint and version
MessageExchange an interactionInput and output contract
TaskRepresent continuing workInterrupted and terminal states
ArtifactReturn a work productValidation and access policy
Security declarationDescribe accepted credentialsTarget-side authorization

What happens during a typical exchange?

Consider an illustrative procurement workflow. A purchasing agent needs a supplier-risk assessment from a separately owned specialist. It resolves an approved Agent Card, checks the endpoint and authentication requirements, and sends a bounded request containing the supplier identifier and necessary context. The specialist can return a direct message for a simple answer or a task for work that continues beyond the initial request.

For a task, the client follows progress through a supported update mechanism. The specialist may need additional input or authorization before continuing. On completion it returns artifacts such as a report or structured assessment. The caller must still evaluate that output before taking a business action. A protocol-level completed state does not mean the result is accurate, approved or suitable for an irreversible transaction.

Source: a2a-protocol.org/latest/specification · Updated 3 November 2026

How should you handle long-running work?

A2A defines task states and operations for retrieving, canceling and following tasks. Implementations can provide polling, streaming or push notifications according to their advertised capabilities. JSON-RPC, gRPC and HTTP+JSON bindings share the semantic model, but their wire formats and streaming behavior differ. Calling A2A simply JSON-RPC over HTTP leaves out the other specified bindings.

Enterprise teams should define timeout, retry and cancellation policies around that lifecycle. Cancellation is not guaranteed rollback of actions already performed. A retried request can have business consequences even when the transport recovered correctly. Record correlation identifiers, test interrupted work and decide who owns an unresolved task after either participant restarts. Those are application and operational decisions, not outcomes to infer from protocol support.

Source: a2a-protocol.org/latest/specification · Updated 3 November 2026

What security checks belong at the boundary?

Agent Cards declare security schemes and requirements. Callers need credentials accepted by the target, and the destination must enforce authorization for the requested work. A signed card can help verify metadata, but a valid signature does not establish that every advertised capability is trustworthy or that the caller may exercise it. Discovery, authentication and authorization should remain separate checks.

Use approved endpoints, limited credentials and an explicit owner for each integration. Treat returned agent content as untrusted input to the next system. Do not place access tokens in task logs or artifacts. Test expired credentials, rejected audiences, missing scopes and cross-user access. A2A does not automatically federate AWS IAM, Microsoft Entra and Google IAM just because all three environments host agents.

Source: a2a-protocol.org/latest/specification , modelcontextprotocol.io/docs/2025-11-25/tutorials/security/security_best_practices · Updated 3 November 2026

Where do MCP, ACP and BAND (band.ai) fit?

MCP exposes tools, resources and prompts to applications. A specialist reached over A2A can use MCP internally to reach its data. Agent Client Protocol connects client applications such as editors to coding agents. These boundaries can coexist, but shared acronyms or JSON-RPC usage do not make their messages interchangeable.

BAND (band.ai) is relevant when the problem extends beyond calling one specialist. Its hosted interaction layer connects agents and humans in shared rooms with @mention routing, memories and activity timelines, while people approve, redirect or override in place. BAND lists A2A, ACP and MCP; its SDK implements an A2A adapter and gateway. This is a product capability to evaluate alongside protocol conformance, not a feature of A2A itself.

Source: band.ai/integrations?utm_source=distributedaiagents.com&utm_medium=referral , band.ai/platform?utm_source=distributedaiagents.com&utm_medium=referral · Updated 3 November 2026

How should you assess an A2A implementation?

Start with one bounded process and two endpoints. Record the supported version, binding, credentials, task lifecycle and artifact format. Exercise a normal request, a request needing clarification, rejected authorization and an interrupted task. Compare observations with the vendor documentation, not a generic A2A checkbox. Version 1.0 changed parts of the contract, so test older interfaces explicitly rather than treating all release families as identical.

The protocols hub summarizes vendor support, while the comparison table separates runtimes from gateways and collaboration layers. Use the RFP checklist to ask for documented behavior and a demonstration. Keep implementation findings separate from this guide's unchanged editorial mesh-fit scores. The decision is whether your particular endpoints interoperate under your policies.

Source: a2a-protocol.org/latest/specification · Updated 3 November 2026

What do enterprise buyers ask?

Is A2A a Google-only protocol?

No. Google introduced A2A, but the specification defines interoperability between independent agent systems, not a Google-only runtime.

Does A2A replace MCP?

No. A2A exchanges work between agents; MCP exposes tools and context. An agent can use both contracts.

Must agents share their internal memory?

No. A2A is designed for independent systems without exposing internal memory, tools or proprietary logic. Applications choose which context to send.

Does every request create a task?

No. A send operation can return a direct message or a task, depending on the interaction and implementation.

Does A2A guarantee business success?

No. Protocol completion and a correct, approved business outcome are different things. Validate artifacts and enforce action policy.

Which related pages should you read?