Skip to content
Search

Agent2Agent

lastmod 2026-01-19created 2025-04-107 edits

An open protocol enabling communication and interoperability between opaque agentic applications.

About

  • A2A is an open protocol designed by Google that complements Anthropic's MCP (Model Context Protocol)
  • Developed to solve the various problems that arise when building large-scale multi-agent systems
  • Provides interoperability by letting agents built by different providers interact
  • Through this protocol, enterprises benefit from being able to manage agents in a unified way across multiple platforms and cloud environments
  • Google Cloud and various partners such as Accenture, Atlassian, Box, and Capgemini are participating

A2A design principles

  • Embrace agentic capabilities: supports natural collaboration between agents even without shared memory, tools, or context
  • Build on existing standards: built on widely used technologies such as HTTP, SSE, and JSON-RPC, so it integrates easily with IT environments
  • Secure by default: provides an OpenAPI-level authentication scheme by default, ensuring enterprise-grade security
  • Support for long-running tasks: supports tasks lasting from seconds to days, with real-time feedback and status updates
  • Modality agnostic: supports not only text but also audio and video streaming, handling various forms of data

How A2A works

  • Interaction between a client agent and a remote agent: the client agent sends task requests, and the remote agent carries them out
  • Capability Discovery: each agent advertises the tasks it can perform through an "Agent Card" in JSON format, making it possible to choose a suitable agent
  • Task management: a Task is a structure defined in A2A that supports both short-running and long-running tasks, and its output is represented as an "artifact"
  • Collaboration: agents exchange messages to share context, replies, artifacts, user instructions, and so on
  • User experience negotiation: message components ("parts") have explicit content types, and the client and remote agent negotiate the UI format (iframe, video, web form, etc.)

Real-world example: automating software engineer hiring

  • When a recruiter enters hiring criteria (location, tech stack, etc.) in the Agentspace interface, an agent searches for suitable candidates
  • It collaborates with other specialized agents to gather candidate information and make recommendations
  • Based on this recommendation list, the user can schedule interviews and then call other agents that perform background checks and the like
  • A2A lets multiple agents integrate to automate the hiring process

The future of agent interoperability

  • A2A enables the development of innovative systems by laying the groundwork for diverse agents to cooperate smoothly
  • Developed as an open protocol that anyone can join and contribute to
  • The full draft specification, sample code, and scenarios are published on GitHub
  • Plans to release a production-ready version with partners later this year

A2A vs MCP

A2A vs MCP summary

  • A2A is designed as a protocol for direct communication and collaboration between agents
  • MCP focuses on LLM-centric context management and integration with external tools

Differences by feature:

  • Purpose:
    • A2A focuses on collaboration between independent agents
    • MCP focuses on a single LLM extending its capabilities with external APIs or tools
  • Architecture:
    • In A2A, different agents exist independently of each other in a client-server structure
    • MCP has an application-LLM-tool structure, with the LLM at the center
  • Communication:
    • A2A is based on web standards such as HTTP, JSON-RPC, and SSE
    • MCP uses JSON-RPC 2.0-based communication and supports HTTP streaming
  • Building blocks:
    • A2A is built from tasks, messages, artifacts, and so on
    • MCP is organized around resources, tools, memory, prompts, and so on
  • Strengths:
    • A2A is strong in asynchronous processing and collaboration
    • MCP is strong in context efficiency, parallel processing, and caching
  • Community:
    • A2A is gaining early support, centered on Google Cloud customers
    • MCP is already widely adopted and has an active developer community

Conclusion

  • A2A and MCP ultimately have similar goals in that both support multi-agent/tool-calling structures for achieving complex goals in AI systems
  • Both protocols lack automatic registration and discovery, so manual configuration is required
  • MCP has first-mover advantage and an active community
  • A2A is growing fast on the strength of Google's strong backing
  • The two protocols are likely to evolve toward standardization through complementarity or integration rather than competition
  • From a developer's point of view, it is a positive trend that provides more choices and open standards

See also

Favorite site

author Zer0 sitemap