[Project Proposal] Agent2Agent Protocol (A2A)

[thisisnotapril]

Project Name

Agent2Agent

Project Description

The Agent2Agent (A2A) Protocol is an open standard designed to facilitate communication and interoperability between independent, potentially opaque AI agent systems. In an ecosystem where agents might be built using different frameworks, languages, or by different vendors, A2A provides a common language and interaction model, enabling agents to discover capabilities, negotiate modalities, and collaborate on complex tasks.

This is especially useful in scenarios where agents do not share tools, skills, or context across perimeters such as company, department, region, or certification requirements.

Alignment with AAIF Mission

The Agent2Agent (A2A) project aligns with the Agentic AI Foundation's (AAIF) mission by providing the interoperable infrastructure necessary to accelerate the adoption of agentic AI.

The mission of the AAIF is to be the place for developers and companies to build trusted shared technology to accelerate agentic AI adoption and to drive the growth of open-source agentic AI in the broader community.

The A2A project supports this mission through the following points of alignment:

• Interoperability and Communication: The A2A Protocol is an open standard designed to facilitate communication and interoperability between independent, heterogeneous AI agent systems. This directly supports the AAIF's goal of building interoperable, open, and community-centric infrastructure.
• Common Language: A2A provides a common language and interaction model, enabling agents built with different frameworks, operating in different security perimeters, or built by different vendors to discover capabilities and negotiate modalities. This addresses the foundational need for agents to operate across platforms, which is a key focus for the AAIF.
• Collaboration on Complex Tasks: By enabling agents to collaborate on complex tasks, A2A helps realize the AAIF's vision of creating a "shared home for tomorrow's intelligent agents".
• Open Standards: As an open standard, A2A aligns with the foundation's commitment to ensuring agentic AI infrastructure evolves openly and predictably at production scale.

In summary, the A2A project provides the technical "connective tissue" required for the autonomous, interoperable AI ecosystems that the AAIF is dedicated to sustaining.

Relation to Existing AAIF Projects

MCP -
The A2A (Agent2Agent) project and the Model Context Protocol (MCP) are primarily viewed as complementary standards that operate at different layers of the AI orchestration stack.

The relationship can be defined by their distinct operational boundaries:

• Vertical vs. Horizontal Integration: MCP is described as a vertical integration layer, acting like a "USB-C port" to connect an AI reasoning engine to external tools, databases, and local data. In contrast, A2A is a horizontal orchestration layer designed for peer-to-peer collaboration, negotiation, and task delegation between independent, autonomous agents. We see adoption increasing across perimeter boundaries where agents cannot share tools, skills, or context. In these cases they communicate via A2A’s application-like interface.
• Internal Tools vs. External Coordination: MCP typically manages an agent's internal tool integration, while A2A handles the external coordination required when agents from different vendors, frameworks, or perimeters must work together.
• Interoperability Efforts: There are active efforts to bridge the protocols. For example, proposals exist to implement A2UI (Agent to User Interface) functionality over MCP, allowing clients to maintain flexibility in using both standards. Additionally, there has been discussion regarding MCP officially supporting A2A tasks as an extension to encourage "single gateway" implementations.

Example Use Cases and Evidence of Adoption

• The Big Three Cloud Providers: We're seeing deep, native integrations across the board. In addition to Google's Gemini Enterprise & the new Gemini Enterprise Agent Platform (the evolution of Vertex AI), AWS has natively integrated A2A protocol support into Amazon Bedrock AgentCore Runtime, and Microsoft is enabling cross-platform agent collaboration in Microsoft 365 Copilot via the new Work IQ API and the Microsoft Agent Framework.
• Enterprise Powerhouses: The SaaS giants are breaking down operational silos. ServiceNow (via its AI Agent Fabric, which embeds A2A natively), Salesforce (via Agentforce and MuleSoft Agent Fabric), and Atlassian (Rovo agents) are using A2A to orchestrate workflows between their CRMs, ITSM systems, and external vendor clouds. SAP has committed to A2A as the primary extensibility layer for SAP Joule, enabling enterprises to plug in code-based agents built with LangGraph or CrewAI directly into the Joule ecosystem. Workday is on the founding partner roster, and Box has shipped its Enhanced Extract Agent in production using A2A to call Gemini for structured-data extraction. On the automation side, UiPath made A2A central to its agentic-automation platform launch, bridging RPA's installed base into the new agent ecosystem.
• Unifying the Standards: Fragmentation is the enemy of adoption. Recognizing this, IBM made the strategic decision to deprecate their competing Agent Communication Protocol (ACP) for the BeeAI platform, officially merging it into the A2A standard under the Linux Foundation. IBM also offers ContextForge, an open-source AI gateway and registry that federates MCP, A2A, and REST/gRPC services behind a single governed endpoint with centralized discovery and guardrails.
• Google ADK (Agent Development Kit): Native A2A support in the ADK running anywhere, especially on Agent Platform Runtime (the evolution of Vertex AI Agent Engine), with first-class RemoteA2aAgent and to_a2a() helpers.
• LangGraph + LangSmith Deployment: Every deployed LangGraph agent gets a hosted A2A endpoint at /a2a/{assistant_id} with an auto-generated Agent Card, streaming via SSE, and LangSmith distributed tracing across multi-agent A2A conversations.
• CrewAI + AMP: Production A2A hosting on the CrewAI AMP platform — A2AServerConfig provisions distributed state, enterprise auth (OIDC, OAuth2, mTLS), REST/JSON-RPC/gRPC transports, and signed webhook push notifications.
• AG2 (formerly AutoGen): Expose or connect agent swarms via A2A; the AG2 repo is tagged with a2a as a core capability.
• Microsoft Agent Framework: Unified A2A integration in both Python and .NET, with hosting samples for Azure Functions, Durable Tasks, and A2A.
• AWS Strands: Model-agnostic open-source agent harness from AWS that wraps agents as network-accessible A2A services.
• Claude Agent SDK: Pair Claude's native MCP tool usage with any A2A SDK to delegate work to remote A2A peers.
• OpenAI Agents SDK: Treat remote A2A-compliant agents directly as callable tools.
• Pydantic AI: Native support via the FastA2A library, using .to_a2a() to convert any Pydantic AI agent into an ASGI A2A application in two lines of code.
• Koog: Full A2A support (server, client, and Koog integration) for the JetBrains JVM agent framework ecosystem.
• Embabel: Rod Johnson's JVM-based framework with a dedicated embabel-agent-a2a Spring module for type-safe Goal-Oriented Action Planning (GOAP) orchestration over A2A.
• NVIDIA NeMo Agent Toolkit: NVIDIA's open-source agent toolkit (v1.6, April 2026) ships first-class A2A server, client, and authentication, and runs natively alongside Google ADK, LangChain, LlamaIndex, CrewAI, and Microsoft Semantic Kernel.
• BeeAI Framework (IBM): The successor to IBM's deprecated ACP, BeeAI now ships native A2AServer and A2AAgent adapters, completing the ACP→A2A migration for the BeeAI platform.
• AI Gateways & Routers: Open-source proxies are emerging to govern, secure, and route agent traffic at enterprise scale, and the leading projects all speak A2A. The Linux Foundation's agentgateway project — donated by Solo.io with contributions from Microsoft, AWS, Cisco, Salesforce, Apple, Adobe, Alibaba, Huawei, and Amdocs — is a Rust-native proxy purpose-built for agentic protocols, with first-class support for A2A, MCP, and LLM routing through a single control point. The CNCF Envoy AI Gateway, backed by Tetrate, Bloomberg, Tencent, and others, brings the same idea to the proven Envoy proxy — Envoy now natively deframes MCP, A2A, and OpenAI-style protocols at the network layer to enforce zero-trust policies on agent traffic (Google Cloud, April 2026). LiteLLM (BerriAI) — already used by Stripe, Netflix, Google ADK, OpenHands, and Greptile — has unified A2A, MCP, and 100+ LLMs behind one OpenAI-compatible gateway, with a dedicated /a2a endpoint and pre-built A2A providers for LangGraph, Vertex AI Agent Engine, Azure AI Foundry, Bedrock AgentCore, and Pydantic AI. And for teams that want event-driven topologies instead of synchronous proxies, Solace Agent Mesh is an open-source agent mesh built around an event broker, designed to orchestrate A2A-compliant agents asynchronously across the enterprise.
• Data Management & Streaming: Reltio has launched AgentFlow™ with native A2A support to access real-time enterprise data, and Confluent has integrated A2A directly into Flink for event-driven streaming agents over replayable Kafka topics — currently in Open Preview.
• Agentic Commerce: This is where the autonomous economy becomes real. The Agent Payments Protocol (AP2) and the Universal Commerce Protocol (UCP) — co-developed by Google with Shopify, Etsy, Walmart, Target, and Wayfair — both leverage A2A for transactional integrity, allowing AI agents to securely negotiate and execute purchases. We're also seeing widespread adoption of x402, the new internet-native standard for "pay-per-request" agent interactions, and Skyfire, which provides the verified economic identity (KYA — "Know Your Agent") and programmable wallets required for agents to act as independent economic actors.
• Identity for Agents: A2A only works if agents can prove who they are. Auth0 (Okta) is co-developing the A2A authentication and authorization spec with Google Cloud, including OAuth/OIDC patterns for agent-to-agent calls and CIBA flows for headless agents — with a working demo merged into the official a2a-samples repo.
• Observability: As agents start to delegate to other agents, end-to-end tracing becomes table stakes. Arize AI was a launch partner for A2A and brings OpenTelemetry-compliant tracing across agent boundaries, so you can debug a multi-agent workflow the same way you debug a microservice mesh.
• Foundation Labs: Cohere is the foundation lab on the A2A launch-partner roster, with a particular focus on air-gapped enterprise A2A deployments for regulated industries — a use case the public-cloud-only labs can't easily serve.
• Telecom: Huawei announced an open-source A2A-T (Agent-to-Agent for Telecom) reference implementation at MWC 2026, jointly released by global telecom partners at TM Forum Accelerate Week in February 2026. The open-source release includes a protocol SDK, a registry/addressing/skill-management center, and a low-code orchestration center — bringing carrier-grade A2A profiles to operator networks.
• Secure Messaging: A2A is designed to support any security and identity setup, using the best existing techniques like OAuth, MTLS, JWTs, and SPIFFE. On top of that, teams have been designing secure messaging conventions on top of the protocol. Cisco spearheaded the AGNTCY project, compiling A2A over SLIM (Secure Low-Latency Interactive Messaging) to guarantee quantum-safe, end-to-end Message Layer Security (MLS) for autonomous systems in highly regulated sectors. NANDA from MIT also provides a full-stack communication plan and uses A2A as the agent transport layer. AP3 from Silence Laboratories extends A2A with privacy-preserving multi-party computation, so agents can collaborate on sensitive data without exposing it.
• Agentic UI: With A2UI and CopilotKit's AG-UI, agents can now seamlessly "speak" frontend interface components to one another. A2UI is a generative UI specification (agents send declarative JSON; clients render with native widgets); AG-UI is the Agent↔User Interaction streaming protocol. Both work over A2A as a transport, and they compose well together.

Technical Committee Sponsor (if identified)

Alan Blount

GitHub Repository URL

https://github.com/a2aproject/A2A

License

Apache 2.0

Governance Model

https://github.com/a2aproject/A2A/blob/main/GOVERNANCE.md

CI/CD & Release Workflow

The core specifications do not have a formal CI/CD process though, there are rigorous gates prior to releasing a new version of this specification. Each of the SDKs, documentation, and tooling separate projects and contain CI/CD automation.

Public-Facing Contribution Process for Specifications

https://github.com/a2aproject/A2A/blob/main/CONTRIBUTING.md

Publicly Accessible Issue Tracker

https://github.com/a2aproject/A2A/issues

External Project Dependencies

None explicitly, but A2A relies on established technologies like JSON-RPC, HTTPS, gRPC, etc. Each SDK and developer tooling will have a minimal footprint of dependencies.

Maintainers & Contributors

https://github.com/a2aproject/A2A/blob/main/MAINTAINERS.md

Leadership Team & Decision Process

https://github.com/a2aproject/A2A/blob/main/GOVERNANCE.md

Roadmap

Stability. Now that the v1 stable spec is out, and most of the SDKs are backwards compatible and it's robberable between v0.3 and v1, the primary objective is stability in spec. Minor versions are in flight but no significant or breaking changes are on the horizon.

Security

https://github.com/a2aproject/A2A/blob/main/SECURITY.md

Website URL

https://a2a-protocol.org/

Documented Governance Practices (if any)

https://github.com/a2aproject/A2A/blob/main/GOVERNANCE.md

Links to Social Media Accounts

No response

Trademark and accounts

• If the project is accepted, I agree to donate all project trademarks and accounts to the AAIF.

Details of Existing Financial Sponsorship

No response

Infrastructure Needs or Requests

No response

Additional Information

No response

Application contact name(s) and email(s)

April Kyle Nassi - anassi@google.com

Contributing or sponsoring entity signatory information

If an organization:

Name	Address	Type (e.g., Delaware corporation)	Signatory name and title	Email address
Google LLC	1600 Amphitheatre Parkway, Mountain View, CA 94043	LLC	April Kyle Nassi, OSPO	anassi@google.com