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 moved under the Linux Foundation's Agentic AI Foundation for vendor-neutral standardization
- 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 has reached GA and entered production operation across multiple cloud platforms
- Like MCP, it is positioned as foundational infrastructure for the agent economy
IMPORTANT
A2A is no longer in "specification consideration" but in "production operation phase." For agent authentication and delegation, design alongside Agent Identity.
The Essence of A2A
A2A's defining characteristic is enabling agent ↔ agent peer communication:
- MCP: AI agent ↔ tools/APIs (master-slave relationship)
- A2A: AI agent ↔ AI agent (peer relationship)
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 | Master-Slave (agent directs) | 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 receiptworking: Processinginput-required: Additional information neededcompleted: Successfailed: 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 (as of May 2026)
Key milestones from the emergence of MCP and A2A to the present.
- November 2024: Anthropic releases MCP
- April 2025: Google announces A2A
- June 2025: A2A project migrates under Linux Foundation's Agentic AI Foundation
- October 2025: OpenID Foundation publishes "Identity Management for Agentic AI" v1.1 (Agent ID systematization)
- December 2025: AGENTS.md donated to Linux Foundation by OpenAI and Anthropic — industry standardization
- April 2026: Microsoft Entra Agent ID GA, Okta for AI Agents GA, A2A v1.0 GA — A2A participation surpasses 150 organizations, production deployment across multiple cloud platforms
- 2026+: MCP + A2A + Agent ID three-layer composition becomes the production-operation standard
Current Status
A2A is no longer in "specification consideration" but in production operation phase. This site covers the 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
- 03-architecture — A2A's position in the three-layer model
Recommended Architecture
Build with ADK, equip with MCP, communicate with A2A, identify with Agent ID
The original three-layer model (Build / Equip / Communicate) has gained Identify — this is the 2026 state of the art. A2A alone cannot handle trust boundaries, so combination with Agent ID is now the production prerequisite.
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 | 03-architecture.md |
| 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: May 2026 Status: Reflects A2A v1.0 GA and Agent ID production-operation phase