Agent2Agent Protocol (A2A)

What is the A2A protocol (Agent2Agent)?

The A2A protocol (Agent2Agent Protocol) is an open communication standard that lets AI agents built by different vendors and frameworks discover each other, delegate tasks and return results without exposing their internal memory, prompts or tools. Each agent publishes an Agent Card that describes what it can do, where to reach it and how to authenticate. A2A was originally developed by Google, which launched it in April 2025.

Key takeaways

• The A2A protocol connects AI agents to other AI agents, while the Model Context Protocol (MCP) connects a single agent to tools and data.
• An A2A Agent Card publicly describes an agent's skills, endpoint and required authentication, and other agents read it before delegating a task.
• The Linux Foundation announced the Agent2Agent protocol project on 23 June 2025, after Google donated the protocol.
• Every A2A delegation adds a hop, so governance teams need to know which agent asked which, on whose behalf and with which scope.

How the A2A protocol works

• Discovery: a remote agent publishes its Agent Card, a JSON document usually served at a well-known URL, listing its skills, endpoint and security schemes.
• Authentication: the client agent obtains credentials outside the protocol, such as an OAuth 2.0 token or an API key, and sends them in HTTP headers.
• Task delegation: the client agent sends a message, typically JSON-RPC 2.0 over HTTPS, and the remote agent handles it as a task with states such as working, completed or failed.
• Results: outputs return as artifacts (text, files or structured data), streamed as they are produced or sent later through push notifications for long-running work.
• Opacity: agents collaborate as black boxes, so neither side has to expose its reasoning, memory or tool access to the other.

A2A or MCP: what does each protocol connect, and what is left to govern?

• MCP: links one agent to tools and data sources, such as a CRM or a database, so the agent acts with the permissions of that connection.
• A2A: links one agent to another agent, often from another vendor, which receives a task and decides itself how to complete it.

The two are complementary: an agent reached through A2A may itself use MCP to call its tools. The governance question is the delegation chain. A2A secures each hop, but a request can travel from an employee to an assistant, then to a supplier's agent, then to its tools. Recording which agent asked which, on whose behalf and with which scope stays with the organizations running the agents. The specification is published on the A2A project site, and a Technical Steering Committee with representatives from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow maintains it.

Examples and use cases

A 300-person logistics company gives its sales team an AI assistant that uses the A2A protocol to delegate prospect research to a data provider's agent and scheduling to a calendar agent. Each agent holds its own tokens. When a sales manager leaves, disabling her account in the identity provider says nothing about credentials issued to agents in her name, and the security lead cannot tell an auditor which external agent received customer data last quarter.

Corma addresses the inventory side of that problem: its shadow AI visibility shows which AI tools teams actually use, access reviews with audit evidence confirm who still holds access to each of them, and AI spend monitoring tracks what they cost.

Related concepts

• AI agent governance
• Non-human identity (NHI)
• Machine identity
• OAuth 2.0
• AI agent governance framework: risks and access control
• AI agent automation with Corma

FAQ

Is A2A the same as MCP?

No. MCP connects an AI agent to tools and data sources, while A2A lets one agent delegate a task to another agent. Many architectures use both.

Who created the A2A protocol?

Google originally developed A2A and launched it in April 2025. The Linux Foundation announced the Agent2Agent protocol project on 23 June 2025, after Google donated the protocol.

What is an Agent Card in A2A?

An Agent Card is a JSON document in which an agent describes its skills, its endpoint and the authentication it requires. Client agents read it to decide whether and how to delegate a task.

How does authentication work in the A2A protocol?

Through standard web mechanisms: the Agent Card declares the accepted schemes, such as OAuth 2.0 or API keys, and the client agent sends its credentials in HTTP headers.

Request a demo