A2A Protocol 1.0 Released, Standardizing Inter-Agent Interaction

The A2A (Agent2Agent) Protocol, led by Google and managed by the Linux Foundation, released v1.0.0 in March 2026, finalizing the standard for interaction between heterogeneous AI agents. A2A is designed to allow agents using different frameworks, such as LangGraph and CrewAI, to delegate tasks and share results without exposing their internal implementations (Opaque Execution).

Relationship with MCP and Scope of Application

A2A is complementary to, not competitive with, MCP (Model Context Protocol). While MCP standardizes tool/API access for a single agent, A2A standardizes discovery, delegation, and result sharing between agents. Therefore, it is unnecessary for calling sub-agents or connecting tools within a single framework, and is suitable for scenarios requiring integration across multiple teams or frameworks.

Key Changes in Version 1.0

v1.0 includes data model changes that are not compatible with v0.3. The key changes are as follows:

• Simplified Data Model: The kind field has been removed, and JSON member names serve as discriminators (e.g., {"kind": "text", "text": "..."} → {"text": "..."}).
• Feature Relocation: supportsExtendedAgentCard has been removed and moved to capabilities.extendedAgentCard.
• Naming Convention Change: JSON serialization has changed from snake_case to camelCase, and some type names have also been modified.
• Version Negotiation: Only Major.Minor versions are sent via the A2A-Version header, and an error is returned if unsupported.

Technical Structure and SDK Support

The protocol consists of three layers: Canonical Data Model, Abstract Operations, and Protocol Binding. The Agent Card is a mandatory JSON document published by the server, through which clients identify the agent's capabilities and skills. Official SDKs for six languages, including Python, JavaScript, Java, C#/.NET, Go, and Rust, are currently provided under the Apache License 2.0.