A2A Protocol: Teaching AI Agents to Collaborate Across Vendors

The Multi-Agent Coordination Problem

Enterprise AI is moving from single-model pipelines to multi-agent systems - networks of specialized AI components that divide complex workflows into manageable subtasks. The vision is clear: an orchestrator agent delegates work to specialist agents, which complete subtasks and return results, with the orchestrator synthesizing the final output.

The problem is coordination. Today, every agent-to-agent interaction requires custom integration code. An orchestrator built with LangGraph cannot natively delegate to a Salesforce agent, a Microsoft Copilot Studio agent, or a third-party specialist model without bespoke connector work. As enterprise agent networks grow, this integration overhead compounds - each new agent added multiplies the integration surface.

What A2A Is

Agent2Agent (A2A) Protocol is Google's open interoperability standard for multi-agent communication, launched in April 2025 and subsequently donated to the Linux Foundation. A2A defines:

• Agent Cards - Standardized JSON documents that describe an agent's capabilities, authentication requirements, and endpoint URLs. Any A2A-compatible system can discover what an agent does by fetching its agent card.
• Task delegation - A standardized request format for assigning work to an agent, with defined status tracking (submitted, working, completed, failed) and support for both synchronous and asynchronous completion.
• HTTP-based transport - A2A uses standard HTTPS, making it compatible with existing enterprise network infrastructure, firewalls and API gateways without specialized networking.

The result is a common protocol layer where agents from different vendors can interact without custom integration between every pair.

Launch Context: April 2025 and the 150+ Organization Ecosystem

A2A launched at Google Cloud Next in April 2025 with 50+ founding partner organizations spanning enterprise software (SAP, ServiceNow, Workday), cloud infrastructure (Atlassian, MongoDB), AI platform vendors (Cohere, Mistral), and system integrators (Accenture, Deloitte, KPMG). By the end of 2025, the ecosystem had grown to 150+ organizations.

The Linux Foundation stewardship - mirroring the governance structure chosen for MCP - signals that A2A is designed as infrastructure rather than a Google product. Vendor-neutral governance removes the adoption barrier of depending on a proprietary standard that a single vendor controls.

A2A and MCP: Complementary Standards, Not Competitors

A2A and MCP (Model Context Protocol) address different coordination problems and are designed to be used together:

MCP handles agent-to-tool connectivity: how a single AI agent connects to external data sources, APIs and services. MCP is the integration layer between an agent and the resources it draws on.

A2A handles agent-to-agent connectivity: how one AI agent delegates work to another AI agent, regardless of vendor. A2A is the coordination layer between agents in a multi-agent system.

In a well-architected enterprise multi-agent system, each individual agent uses MCP to access its data sources and tools, while the agents coordinate with each other using A2A. The two standards form complementary layers of the enterprise AI stack.

Enterprise Workflow Examples

A2A's practical impact is clearest in enterprise workflows where specialized agents from different vendors must collaborate:

Contract review workflow: A document intake agent (Microsoft) extracts contract text and delegates clause analysis to a specialized legal AI agent (third-party), which returns structured findings to a risk scoring agent (internal), which routes to a human reviewer agent for exceptions above a defined threshold. Each agent is from a different vendor; A2A handles the delegation and status tracking at every step.

Customer onboarding workflow: A CRM orchestrator (Salesforce) delegates identity verification to a compliance agent (specialized provider), credit assessment to a financial AI agent, and welcome communications to a content generation agent. Results converge back at the orchestrator without any custom integration between the specialist agents.

Isotropic designs A2A-native multi-agent architectures where agent boundaries, capability definitions, and coordination patterns are engineered from the start - not retrofitted when integration complexity grows. Contact business@isotrp.com to discuss how A2A applies to your multi-agent AI program.