Agent-to-agent (A2A) protocol

Agents should talk like agents

A2A is the open standard that lets AI agents from different companies find each other, sign in securely and hand off work over the internet. Here’s how it works, and where it’s heading for customer-facing agents.

Book a demoHow A2A works
G2 rating: 4.9 out of 5SOC 2 Type II & SOC 3GDPRCCPAHIPAA
0:00

Concept demo. Two agents that meet on a phone call switch to a direct connection to finish a booking. A2A runs over the internet today; there is no standard yet for this phone handoff.

What is the A2A protocol?

One open standard for agents to work together

A2A (Agent2Agent) lets AI agents built by different companies find each other, sign in securely and hand off work. Google introduced it in April 2025. It is now part of the Linux Foundation’s Agentic AI Foundation, supported by more than 150 organizations, and version 1.0, its first stable spec, shipped in March 2026.

A2A and MCP

How A2A compares with MCP
Compared onMCPA2A
ConnectsAn agent to its own tools and dataAn agent to other agents
AddsDepth inside one agentReach across teams and companies
ExampleA mechanic agent reads a diagnostic scannerA repair shop’s agent orders parts from a supplier’s agent
The other sideA tool you controlAn agent that keeps its own memory, tools and prompts private

They work together: MCP gives each agent depth, A2A connects agents to each other.

How A2A works

What happens when two agents connect

Book a demo
  1. 01DiscoverThe client agent reads the other agent’s Agent Card: a public file that lists what it can do and how to connect.
  2. 02AuthenticateIt signs in over HTTPS with a method the card lists, such as OAuth 2.0. The other agent decides what it is allowed to do.
  3. 03SendIt sends the request as a task. A task can carry text, files and structured data.
  4. 04ResolveThe other agent reports progress until the task is completed or needs more input, then returns the result.

Where A2A is today

Already running between agents

Book a demo
  • Agentic AI Foundation · August 2026

    Built into the major clouds

    Google Cloud, Microsoft Azure AI Foundry and AWS Bedrock AgentCore support A2A, so agents built on each can work together.

    AAIF
  • Agentic AI Foundation · August 2026

    Already in production

    Huawei’s Celia assistant uses A2A with in-app agents on HarmonyOS, and Google Cloud and PayPal use it with the Agent Payments Protocol for agent-led shopping.

    AAIF
  • Google · September 2026

    Customers’ AI is reaching businesses

    Gemini is testing calls to local businesses for Pixel 11 owners in the US. For now, assistants like it use phone lines and chat windows built for people.

    TechCrunch

Questions about the A2A protocol

Book a demo

A2A (Agent2Agent) is an open protocol that lets AI agents built by different companies communicate and work together over the internet. Google introduced it in April 2025. It is now governed by the Agentic AI Foundation, part of the Linux Foundation, with more than 150 supporting organizations.

It is one AI agent handing work to another and getting the result back. Without a shared standard, every pair of agents needs its own integration, or they fall back to channels built for people, like chat windows and phone calls. A2A gives agents one way to find each other, sign in and exchange tasks.

MCP connects an agent to its own tools and data. A2A connects an agent to other agents, including ones run by other companies. They work together: most agent systems use MCP inside each agent and A2A between agents.

A short JSON file an agent publishes on its domain at /.well-known/agent-card.json. It lists the agent’s name, what it can do, where to reach it and how to sign in, so other agents can decide whether and how to work with it.

A2A runs over HTTPS and uses standard web sign-in, such as OAuth 2.0, OpenID Connect and API keys, listed in the Agent Card. Each agent decides what a caller is allowed to do, and it shares tasks and results, never its internal memory, tools or prompts.

Not today. A2A runs over the internet. There is no standard yet for two AI agents that meet on a phone call to recognize each other and switch to a direct connection. Our demo video shows one way it could work.

No. Gibberlink was a 2025 hackathon demo in which two voice agents on a call switched to sound-encoded data. A2A is an open standard for agents to connect directly over the internet, with sign-in and task tracking built in.

We scope agent-to-agent connections with each customer. Book a demo, tell us which agents you want yours to work with, and we will walk you through what is possible today.

Plan for agents that work with agents

Book a demo to talk through which agents you’d want yours to work with, or watch the concept demo.

Book a demoWatch the demo