====== Agent Protocol ======
The **Agent Protocol** originated as a community-driven open standard defining a unified HTTP API interface for communicating with AI agents. Initially developed by the AI Engineer Foundation, the concept of standardizing agent-to-agent communication evolved significantly.((AI Engineer Foundation. "Agent Protocol Specification." [[https://[[github|github]].com/AI-Engineer-Foundation/agent-protocol|github.com/AI-Engineer-Foundation/agent-protocol]], 2023.)) By mid-2025, Google launched the **Agent2Agent (A2A) Protocol** under the Linux Foundation's governance, which has become the primary open standard for secure, interoperable communication between AI agents across platforms and vendors.((Google. "Agent2Agent (A2A) Protocol." [[https://[[github|github]].com/a2aproject/A2A|github.com/a2aproject/A2A]], 2025.))

===== Original Agent Protocol =====
The original Agent Protocol specification defined common REST API endpoints for interacting with any AI agent:

  * **POST /ap/v1/agent/tasks** - Create a new task
  * **GET /ap/v1/agent/tasks/{id}** - Retrieve task status
  * **POST /ap/v1/agent/tasks/{id}/steps** - Execute the next step
  * **GET /ap/v1/agent/tasks/{id}/artifacts** - List task artifacts

This standardization enabled framework-agnostic agent interaction. Frameworks like [[autogpt|AutoGPT]] and [[babyagi|BabyAGI]] adopted it to ensure agents built on different stacks could communicate through a common interface. The specification used JSON payloads over HTTP with clear lifecycle management for tasks and steps.

===== Agent2Agent (A2A) Protocol =====
The **[[a2a_protocol|A2A Protocol]]**, launched in June 2025 by [[google|Google]] and governed by the Linux Foundation, extends the vision of agent interoperability to enterprise scale:

  * **[[github|GitHub]]:** [[https://github.com/a2aproject/A2A|github.com/a2aproject/A2A]]
  * **Standards:** Built on HTTP, SSE (Server-Sent Events), and JSON-RPC
  * **Governance:** Linux Foundation with vendor-neutral oversight

=== Key Capabilities ===
  * **Agent Discovery:** Agents can find and identify each other's capabilities dynamically
  * **Secure Messaging:** Enterprise-grade authentication with OpenAPI parity, authorization controls
  * **Task Coordination:** Support for long-running tasks with real-time feedback, notifications, and state updates
  * **[[human_in_the_loop|Human-in-the-Loop]]:** Built-in support for processes requiring human approval, lasting hours or days

=== Adoption ===
Over 100 technology companies are involved, including:

  * **[[google|Google]] Cloud** - Core developer
  * **[[microsoft|Microsoft]]** (Azure AI Foundry) - Includes A2A alongside [[semantic_kernel|Semantic Kernel]] and [[autogen|AutoGen]]
  * **[[salesforce|Salesforce]]** ([[agentforce|Agentforce]]) - Cross-platform orchestration
  * **AWS, Cisco (Outshift/AGNTCY), SAP, [[servicenow|ServiceNow]]**

=== Relationship to MCP ===
A2A and [[anthropic_context_protocol|MCP]] are complementary standards:

  * **MCP** standardizes how agents connect to tools and data sources (agent-to-tool)
  * **A2A** standardizes how agents communicate with each other (agent-to-agent)

Together they form the emerging interoperability stack for [[agentic_ai|agentic AI]] systems. Bain's 2025 analysis projects 5-10% of enterprise tech spending directed toward agent platforms using these standards over the next 3-5 years.

===== See Also =====
  * [[agent_network_protocol|Agent Network Protocol]]
  * [[a2a_vs_mcp_vs_agui|A2A vs MCP vs AG-UI]]
  * [[ag_ui_protocol|AG-UI Protocol]]
  * [[agent_communication_protocol|Agent Communication Protocol]]
  * [[agent_interface_design|Agent Interface Design]]

===== References =====