A2A (Agent-to-Agent Protocol): What is it?

An open standard protocol enabling agents to collaborate and delegate tasks over networks

About This Document

This document explains A2A's core concepts, key differences from MCP, benefits and challenges, and future outlook. For detailed comparison with sub-agents, also refer to what-is-subagent.md.

For the overall taxonomy of agent terminology (architecture patterns, execution roles, implementation units), see agent-taxonomy.md.

What is A2A?

A2A (Agent-to-Agent Protocol) is an open standard protocol that enables different AI agents to communicate and collaborate in peer-to-peer relationships over networks.

Background and Leadership

Here is a brief history of A2A's origins and standardization.

• Google led the effort, announcing A2A in April 2025
• In June 2025, A2A was donated to the Linux Foundation's Agent2Agent project for vendor-neutral stewardship
• As of April 2026, over 150 organizations participate in the A2A protocol project (AWS, Cisco, Google, IBM, Microsoft, Salesforce, SAP, ServiceNow, and others)
• A2A v1.0 is a stable release, with adoption advancing across multiple cloud platforms
• Around August 2026, moves to place A2A alongside MCP under the Agentic AI Foundation (AAIF) were reported (confirm governance details at the time of reading)
• Like MCP, it is positioned as one foundational protocol for agent collaboration

IMPORTANT

A2A is past pure specification work: v1.0 and cloud implementations are in use. That does not mean adoption alone is enough. Design authentication and delegation with Agent Identity.

The Essence of A2A

A2A's defining characteristic is enabling agent ↔ agent peer communication:

• MCP: AI agent ↔ tools/APIs (the agent calls tools)
• A2A: AI agent ↔ AI agent (either side can request work)

In One Sentence

"A protocol allowing different AI agents to request work from each other"

Architectural Model

Three-layer structure for agent development:

• Build with ADK: Construct the agent itself
• Equip with MCP: Connect tools and APIs
• Communicate with A2A: Communicate with other agents

Why A2A?

Current Challenges

As AI agent technology evolves rapidly, companies and organizations develop and operate their own agents. However, there is no standard communication mechanism between these agents.

The Silo Problem:

• Internal Sales Analysis Agent → wants to query Salesforce AI Agent
• But there's no standard communication protocol
• MCP enables "agent ↔ tools" but not "agent ↔ agent"
• Result: agents operate in isolation, making cross-organization collaboration difficult

Before and After A2A

Before A2A:

• Agent-to-agent communication handled by custom implementations (API integration, etc.)
• Without standards, each integration requires custom specification negotiation
• Not scalable

After A2A:

• Standardized protocol enables agent-to-agent communication
• Authentication, authorization, and trust models are unified
• Agents from different organizations can collaborate seamlessly

Communication Flow

The diagram below illustrates the basic flow of agents from different organizations communicating via the A2A protocol.

Fundamental Differences from MCP

While both A2A and MCP enable "connections," their connection targets are fundamentally different. The comparison table below clarifies the distinctions.

Feature Comparison

The following table contrasts MCP and A2A.

Aspect	MCP	A2A
Led by	Anthropic	Google → Linux Foundation
Purpose	Agent ↔ Tools	Agent ↔ Agent
Connection Target	MCP servers (you manage)	Other agents (including external parties)
Communication Model	Client-driven (agent calls tools)	Peer-to-peer (both can request)
Context Sharing	Shareable with parent agent	Fully isolated (expects opaque parties)
Owner	Self	Self or others
Trust Model	Implicit trust	Authentication/authorization required

Decision Flowchart: Which Should You Choose?

Use the following flowchart to determine the right choice among MCP, A2A, and custom sub-agents.

Core Concepts of A2A

Three key concepts are essential in A2A.

1. Agent Card

An agent's "self-introduction card" that provides information helping other agents and discovery systems understand what capabilities an agent offers.

Format: JSON

Included Information:

• Agent name and description
• Supported Capabilities
• Authentication method
• Supported A2A versions
• Endpoint information

Location: Discoverable via /.well-known/agent.json

2. Task

A "unit of work" agents request from each other. Handles both brief interactions and extended processes.

Lifecycle:

• submitted: Immediately after task receipt
• working: Processing
• input-required: Additional information needed
• completed: Success
• failed: Failure

Characteristics:

• Asynchronous execution
• Long-running task support
• Polling or callback mechanisms

3. Artifact

Data generated as a task execution result. Not limited to text—includes files and structured data.

Supported Formats:

• Text (plain text, markdown, etc.)
• Files (PDF, images, etc.)
• Structured data (JSON, XML, etc.)
• Multimodal (images, audio anticipated)

Benefits of A2A

Adopting A2A provides the following advantages.

✅ Standardized Cross-Organization Collaboration

Agents from different organizations can communicate via a standard protocol. This marks a shift from "custom per-company integration" to "standards-based collaboration."

✅ Agent Specialization and Cooperation

Each agent can specialize in its domain, while complex tasks benefit from multi-agent collaboration. For example:

• Sales Agent ← retrieves forecast data from Market Analysis Agent
• Sales Agent ← retrieves customer info from CRM Agent

✅ Multimodal Support

Information exchange isn't limited to text—images, audio, and files enable richer communication.

✅ Asynchronous Task Support

Long-running operations like report generation can execute in the background, with results retrieved later.

✅ Vendor Independence

As an open standard under Linux Foundation stewardship, there's no vendor lock-in.

Drawbacks and Challenges

A2A also comes with challenges that remain to be addressed.

❌ Implementation Complexity

Authentication, authorization, and encryption are mandatory—requirements stricter than MCP.

❌ Network Dependency

Requires handling network latency, connection dropouts, and other failures. Timeout configuration and retry logic are essential.

❌ Debugging Difficulty

Tracing inter-agent communication is complex, making troubleshooting challenging. Logging strategy is critical.

❌ Maturity is Still Low

Compared to MCP (released November 2024), A2A is even newer. Implementation patterns and operational best practices remain limited.

❌ Nascent Ecosystem

Only a limited number of agents and tools support A2A yet. Broader adoption will take time.

❌ Trust Challenges

You cannot fully guarantee the quality or security of agents owned by others. Risks of misuse by malicious agents must be considered.

Choosing Between Sub-agents and A2A

For collaboration between agents you manage, use "custom sub-agents." For external agent collaboration including other organizations, use "A2A."

Comparison

The following table contrasts custom sub-agents with A2A agents.

Dimension	Custom Sub-agent	A2A Agent
Location	Within same process	Over the network
Owner	Self	Self or others
Trust Relationship	Full trust	Authentication/authorization required
Context	Partially shared with parent	Fully isolated
Lifecycle	Session-scoped	Persistent service

Understanding via Metaphor

• Sub-agents = "In-house specialist department"

  • Located in the same building (process)
  • Under supervisor (parent agent) oversight
  • Fully trustworthy

• A2A agents = "Outsourced partner / vendor"

  • Located at separate facilities (separate processes)
  • Communicate over networks
  • Relationship based on contracts (authentication/authorization)

Complementary, Not Competing

Sub-agents and A2A agents don't compete—they complement each other. Most systems use both:

• Internal task decomposition → Custom sub-agents
• External resource utilization → A2A agents

Current Maturity and Future Outlook

Timeline (approx. as of August 2026)

Key milestones from MCP and A2A to the present. Dates follow public announcements and press; confirm governance names at the time of reading.

• November 2024: Anthropic releases MCP
• April 2025: Google announces A2A
• June 2025: A2A is donated to the Linux Foundation's Agent2Agent project
• October 2025: OpenID Foundation publishes "Identity Management for Agentic AI" v1.1 (Agent ID systematization)
• December 2025: Moves such as AGENTS.md contributions into Linux Foundation–related frameworks
• March–April 2026: A2A v1.0, reports of 150+ supporting organizations, and cloud implementations advance; product announcements such as Microsoft Entra Agent ID and Okta for AI Agents continue
• Around August 2026: Reports of moving A2A toward the Agentic AI Foundation (AAIF)
• Design combination: MCP (tool connection) + A2A (agent-to-agent) + Agent ID (identity and delegation) is one stack often seen in production. It is not another name for the five-layer model

Current Status

A2A is past pure specification work: a stable release and implementations are advancing. This book covers supporting elements on these pages:

• Agent Identity — Identification and delegation underpinning A2A communication
• Multi-Agent / Agent Teams — Evolution from intra-org Agent Teams to cross-org Agent Mesh
• II.1 Five layers — A2A's position among the five layers

Recommended Architecture

Build with ADK, equip with MCP, communicate with A2A, identify with Agent ID

The original three-layer model (Build / Equip / Communicate) plus Identify is a common 2026 framing. A2A alone cannot handle trust boundaries. For cross-organization traffic, design it together with Agent ID.

Remaining Open Issues

A2A has reached GA, but the ecosystem still has unresolved areas:

• Cross-vendor interoperability — Inter-vendor Agent ID interop is not yet established (OIDC-A, IPSIE, AuthZEN standardization is in progress)
• Scope attenuation in recursive delegation — Design needed so authority does not balloon across A2A-mediated sub-agent chains (OAuth Token Exchange, Macaroons, etc.)
• Revocation propagation — Mechanism for revocation in one place to reach all endpoints (Shared Signals Framework, etc.)
• Authentication for browser-controlling agents — Web Bot Auth draft in progress

Details in Agent Identity / Remaining Issues.

What to Read Next

Explore these documents to deepen your A2A understanding:

Purpose	Document
Identification and delegation in A2A	Agent Identity
Relation to intra-org Agent Teams	Multi-Agent / Agent Teams
Sub-agent details	what-is-subagent.md
Agent terminology organization	Agent Taxonomy
MCP details	what-is-mcp.md
Overall architecture	II.1 Five layers
About Skills	what-is-skills.md

Sources

• A2A Protocol Surpasses 150 Organizations (Linux Foundation, April 2026) — 150 organizations, production deployment
• OpenID Foundation — Identity Management for Agentic AI v1.1 — Agent ID systematization
• AGENTS.md official site — Linux Foundation standard

Last Updated: August 2026 Status: Governance wording corrected; claims softened; distinguished from the five-layer model