Google Announces the Agent2Agent Protocol
Google announced A2A on April 9, 2025, with more than 50 partners. We explain Agent Cards, task lifecycles, and artifacts, distinguish A2A from MCP, and show which technical and governance questions the draft deliberately leaves open.
Agents That Cannot Talk to Each Other
Enterprises are beginning to operate several AI agents at once: a procurement agent on LangGraph, a support agent on CrewAI, a vendor agent inside Salesforce or SAP. Each framework brings its own message formats, state model, and calling conventions. Agents from different stacks therefore cannot delegate work without custom integration code. Every new pairing remains bespoke, and the number of connections grows quadratically with the number of agents.
One half of this problem already has a standard. Since November 2024, Anthropic's Model Context Protocol (MCP) has described how an agent connects to tools and data. The other half — communication between agents — lacked an equivalent contract. Point-to-point bridges exist, but they do not solve interoperability. A2A is intended to close precisely that gap.
What Google Announced on April 9
On April 9, 2025, at Cloud Next in Las Vegas, Google announced the Agent2Agent protocol (A2A): an open protocol for communication between AI agents, launched with support from more than 50 technology partners — among them Atlassian, Box, Cohere, LangChain, MongoDB, PayPal, Salesforce, SAP, ServiceNow, and Workday — plus service providers including Accenture, Deloitte, and McKinsey.
The specification is a draft, published as open source on GitHub, announced alongside Google's open-source Agent Development Kit (ADK) and an AI Agent Marketplace. A2A builds on HTTP, Server-Sent Events, and JSON-RPC — deliberately unspectacular standards that pass through enterprise firewalls. The model defines two roles: a client agent formulates tasks; a remote agent executes them. Google states that a production-ready version is planned for later in 2025.
Agent Cards and Capability Discovery
Discovery works through the Agent Card: a JSON document that a remote agent serves at a well-known URL, typically /.well-known/agent.json. The card declares identity, skills, endpoint, supported input and output modalities, and authentication requirements. A client agent reads the card and decides whether this counterpart can handle a given task — machine-readable capability advertisement, comparable to what OpenAPI descriptions did for REST APIs.
That is where discovery ends in the current draft. It defines no registry, search infrastructure, or vetting mechanism. An Agent Card describes a known candidate; it does not explain how a client finds that candidate or why the client should trust it. Both questions remain outside the specification.
Tasks as the Unit of Work
A2A communication is oriented around tasks. A task is a stateful object with a unique ID and a defined lifecycle: submitted, working, input-required, completed, failed, canceled. Outputs are artifacts, composed of parts with negotiated content types — text, files, structured data. The protocol is modality-agnostic by design; audio and video are named as goals.
Long-running work is a core case, not an edge case. Status updates stream over SSE, push notifications cover tasks that run for hours or days, and the input-required state models human-in-the-loop steps. This is the central design decision: A2A treats inter-agent work as asynchronous jobs with observable state, not as a chat transcript.
A2A and MCP Are Complements
Google positions A2A explicitly as a complement to Anthropic's MCP, announced on November 25, 2024. Amin Vahdat, Google Cloud's VP for ML, Systems and Cloud AI, summarized the split at launch: MCP connects models to data and tools; A2A provides interoperability between agents. In practice an agent will use both — MCP downward to its tools, A2A sideways to its peers.
This division is technically meaningful. An MCP tool is called with defined inputs and returns a result. Work is delegated to an A2A remote agent instead: it plans independently, asks follow-up questions, and may refuse the assignment. The protocols therefore encode different contracts — function-call semantics on one side, task lifecycles on the other.
| Aspect | MCP (Anthropic) | A2A (Google) |
|---|---|---|
| Announced | November 25, 2024 | April 9, 2025 |
| Connects | Agent to tools and data | Agent to agent |
| Interface unit | Tools, resources, prompts | Tasks with artifacts |
| Counterpart | Deterministic tool call | Opaque autonomous agent |
| Discovery | Configured servers | Agent Card at a well-known URL |
What A2A Does Not Do
A2A does not share memory, context windows, or tools between agents. Agents remain deliberately opaque to each other; only tasks, messages, and artifacts cross the boundary. A2A also does not orchestrate: which agent receives which task, and in what order, remains the application's decision.
Semantic interoperability does not follow automatically either: two agents may both speak A2A and still misread each other's artifacts. The draft status adds another constraint. Authentication follows OpenAPI practice, but cross-organizational identity, agent registries, and liability for delegated actions remain open. Production use in April 2025 therefore means building on a moving target. We use A2A as an interface for prototypes, not yet as an immutable foundation.
Open Governance as a Signal
The partner list is itself a governance statement. Salesforce, SAP, ServiceNow, and Workday compete with Google in several markets; their names at launch signal that A2A is meant as an industry protocol, not a Google API. Google says it will develop the protocol in the open with clear contribution pathways, and has released the draft under an open-source license.
History teaches caution and hope in equal measure. Protocols gain value through neutrality; standards controlled by a single vendor tend to fragment. The credible next step would be transferring A2A to a neutral foundation with shared governance. Whether Google takes that step will say more than any detail of the specification.
The Outlook From April 2025
From the vantage point of April 2025, the stack is taking shape as two thin layers: MCP for agent-to-tool and A2A for agent-to-agent. That division will hold only if a production-ready version arrives this year and independent implementations follow. We are therefore watching three areas: registry and discovery infrastructure, verifiable agent identity, and the response from Anthropic and OpenAI.
The practical consequence does not depend on the outcome. Design agent boundaries as task interfaces today, with explicit inputs, observable state, and typed artifacts. That structure maps to A2A even if the final specification changes in detail. Within twelve months, the protocol will either become the interoperability layer for enterprise agents or splinter into vendor dialects. More than 50 launch partners are a strong signal for the first path, not yet proof.
Sources
- Announcing the Agent2Agent Protocol (A2A) — Google Developers Blog (April 9, 2025)
- Build and manage multi-system agents with Vertex AI — Google Cloud Blog (April 9, 2025)
- Google's Agent2Agent Protocol Helps AI Agents Talk to Each Other — The New Stack (April 9, 2025)
- Introducing the Model Context Protocol — Anthropic (November 25, 2024)
- A2A protocol draft specification — GitHub google/A2A (April 2025)