Today we're announcing the Agent2Agent (A2A) Protocol — an open standard that enables AI agents from different vendors and frameworks to communicate, coordinate, and collaborate on tasks.
Alongside more than 50 technology partners including Atlassian, Box, Cohere, Intuit, LinkedIn, MongoDB, Salesforce, SAP, ServiceNow, UKG, and Workday, Google is contributing A2A to the broader AI ecosystem as an open protocol. This is the infrastructure that makes multi-vendor agentic AI practical.
The Problem A2A Solves
AI agents are becoming the primary way enterprises automate work. But today, most agents operate in silos. A recruiting agent built on one platform can't delegate a task to a screening agent built on another. An orchestrating agent from one vendor can't discover, invoke, or monitor agents from a different vendor's stack.
A2A fixes this. It's a common language for agents to advertise their capabilities, accept tasks, exchange state, and stream results — regardless of the underlying model, framework, or vendor.
A2A is to multi-agent systems what HTTP is to the web: a transport protocol that makes the ecosystem composable.
Five Design Principles
- Embrace agentic capabilities. A2A is designed for agents that have memory, tools, and the ability to take long-running actions — not just stateless request-response APIs.
- Built on existing standards. A2A uses HTTP, JSON-RPC, and Server-Sent Events. No new transport primitives. Any stack that speaks HTTP can participate.
- Secure by default. Authentication and authorization are first-class. Agents authenticate before accepting tasks, and enterprise security policies flow through the protocol.
- Supports long-running tasks. Agents can pause for human input, report progress over time, and resume after interruption. The protocol models task lifecycle, not just completion.
- Modality agnostic. A2A handles text, structured data, files, forms, and streaming output. The same protocol works for a code-writing agent and a document-processing agent.
How It Works
Agent Cards
Each A2A-compliant agent publishes an Agent Card — a JSON document that describes what the agent can do, what inputs it accepts, what outputs it produces, and where to reach it. Agent Cards are the discovery layer: an orchestrator queries Agent Cards to find the right sub-agent for a task.
Tasks
The core unit of work in A2A is a Task. Tasks move through a defined lifecycle:
submitted→ the orchestrator sends a task to a remote agentworking→ the agent is executinginput-required→ the agent needs clarification or human approvalcompleted/failed/cancelled→ terminal states
Agents can push streaming updates during the working state via Server-Sent Events, so orchestrators get live progress rather than a polling loop.
Example: Candidate Sourcing
A hiring manager asks an orchestrating agent to find qualified software engineers in Berlin. The orchestrator uses Agent Cards to discover three specialist agents: one that searches LinkedIn, one that screens resumes against the job description, and one that schedules outreach.
The orchestrator delegates each sub-task via A2A. Each sub-agent reports progress. The screener hits an ambiguity — is remote work acceptable? — and transitions the task to input-required. The orchestrator routes that question to the hiring manager, gets a response, and resumes the task. The final output — a shortlist with interview slots already booked — flows back to the orchestrator, composed from three independent agents across two vendors.
None of this required the agents to share a framework, a model provider, or a deployment platform. A2A is the glue.
A2A and MCP: Complementary, Not Competing
Anthropic's Model Context Protocol (MCP) connects agents to tools and data sources: databases, APIs, file systems, code interpreters. A2A connects agents to other agents. They solve different problems in the same stack.
A well-architected multi-agent system will use both: MCP for tool and resource access within an agent, A2A for task delegation between agents. The two protocols are designed to compose cleanly.
What Comes Next
The A2A specification and reference implementation are open source on GitHub under the Google DeepMind organization. The protocol is in active development, and the 50+ launch partners are already building production integrations.
If you're building agentic systems and need your agents to collaborate across vendor boundaries, A2A is the infrastructure layer to watch. The era of siloed agents is ending.