Agent-to-Agent Protocol (A2A)

The Agent-to-Agent protocol (A2A) is an open standard that defines how AI agents communicate with, discover, and delegate tasks to one another. Where earlier AI integration standards focused on connecting agents to tools and data, A2A addresses a different problem: how agents talk to each other.

As AI agents proliferate across organisations — and across the products and services organisations rely on — the ability for those agents to collaborate becomes critical. Without a shared protocol, every agent-to-agent interaction requires bespoke integration: a fragile, expensive, and unscalable approach. A2A provides the interoperability layer that makes multi-agent systems practical.

In practical terms, A2A allows a general-purpose orchestrating agent to discover what a specialist agent can do, hand off a well-defined task, receive results, and incorporate them into a larger workflow — all without either agent having been purpose-built for the other.

Google introduced A2A in April 2025 as an open-source specification, with a founding group of over fifty technology partners spanning cloud providers, enterprise software vendors, and AI infrastructure companies. Like MCP before it, A2A was released as an open standard from day one — a signal that the industry recognised agent interoperability as a problem requiring a shared solution, not a competitive moat.

The timing reflected the maturity of the AI industry. By early 2025, organisations were moving beyond experimenting with individual AI tools and beginning to deploy multiple agents that needed to work together. The absence of a standard protocol was already creating friction: agents built on different platforms could not communicate without custom middleware, and the complexity compounded with each new agent added to the ecosystem.

A2A drew on lessons from MCP's rapid adoption — particularly the importance of keeping the protocol simple, transport-agnostic, and grounded in existing web standards. Adoption across major AI platforms accelerated quickly, and A2A is now considered a foundational standard alongside MCP for enterprise AI architectures.

A2A is built on three core concepts: discovery, task delegation, and capability negotiation.

Agent Cards

Every A2A-compatible agent publishes an Agent Card — a structured JSON document describing what the agent can do, how to reach it, what authentication it requires, and what data formats it supports. Agent Cards are the discovery mechanism of the protocol: an orchestrating agent reads another agent's card to understand whether it is the right agent for a task, before any task handoff occurs.

Task Lifecycle

Tasks are the unit of work in A2A. A client agent submits a task to a server agent, which acknowledges receipt, executes the work, and returns a result. The protocol defines a clear state machine for tasks — submitted, working, completed, failed — and supports both synchronous responses and asynchronous patterns including streaming and push notifications. This makes A2A suitable for both quick lookups and long-running agentic operations.

Capability Negotiation

Before delegating a task, agents can negotiate over supported input and output modalities. A2A is designed to handle not just text but also structured data, files, and other content types — reflecting the reality that enterprise agents often work with documents, images, and structured records rather than plain text alone.

Authentication is handled via OAuth 2.0 and standard HTTP security patterns, making A2A compatible with enterprise identity and access management infrastructure. All communication occurs over HTTPS, and the protocol is designed to work behind standard API gateways and service meshes.

A2A and MCP solve different problems and are designed to be used together. Understanding the distinction is essential for architecting multi-agent systems.

MCP defines how an agent connects to tools, data, and services — the interaction between an agent and the systems it acts upon. It is the protocol for agent-to-environment communication.

A2A defines how agents communicate with each other — the interaction between agents that collaborate, delegate, and coordinate. It is the protocol for agent-to-agent communication.

In a well-architected multi-agent system, both protocols play a role. An orchestrating agent uses A2A to discover and delegate to specialist agents. Each specialist agent uses MCP to access the tools and data it needs to complete its assigned task. The two protocols form complementary layers of the same stack: MCP as the capability layer, A2A as the coordination layer.

This layered approach is analogous to how the internet itself works: different protocols handle different concerns at different layers, and the combination is more powerful than any single protocol could be.

Multi-Agent Orchestration

A central orchestrating agent receives a complex request and breaks it into subtasks, delegating each to a specialist agent best suited for it — a research agent, a writing agent, a data analysis agent, a compliance review agent. A2A provides the protocol that makes this delegation reliable and vendor-neutral.

Cross-Platform Agent Collaboration

Organisations that deploy agents from multiple vendors — a Microsoft Copilot for productivity, a specialist HR agent, a customer service agent from a niche provider — can integrate them into coherent workflows using A2A, without requiring those vendors to build proprietary integrations with each other.

Delegating to Specialist AI Systems

Some tasks require deep domain expertise that a general-purpose agent lacks. A2A allows a generalist agent to recognise its limitations and route the task to a specialist — a medical coding agent, a legal review agent, an engineering calculation agent — and incorporate the result seamlessly.

Long-Running Agentic Workflows

A2A's support for asynchronous task patterns and push notifications makes it well-suited to workflows that unfold over minutes or hours: research tasks, document processing pipelines, approval workflows, and multi-step data transformation processes.

The most significant AI gains in the coming years will come not from individual AI tools but from coordinated networks of agents working together. Organisations that architect for agent interoperability today are building the foundation for that future. Those that do not will find themselves managing a fragmented collection of siloed AI capabilities that cannot multiply each other's value.

Adopting A2A as a standard yields concrete strategic benefits. It preserves vendor flexibility — organisations can choose the best agent for each task without being locked into a single platform's ecosystem. It reduces integration costs as new agents are added — each new A2A-compatible agent is immediately reachable by all existing orchestrators. And it creates a reusable platform that grows more valuable as the ecosystem around it expands.

A2A also has governance implications. When agent-to-agent communication flows through a standard protocol layer, organisations can instrument that layer: logging what agents delegated what tasks, monitoring outcomes, enforcing policies on which agents can reach which others. This visibility is essential as AI systems take on more consequential roles.

For organisations building an AI Front Door — a governed, discoverable gateway to their AI capabilities — A2A is what allows that gateway to serve not just human users but the growing population of AI agents that will need to interact with it.

Machine Experience (MX) — the design discipline focused on how machines interact with systems and services — does not end at the boundary between agent and tool. As the AI ecosystem matures, machines increasingly need to interact not just with databases and APIs but with other machines: agents that have their own capabilities, interfaces, and expectations.

A2A is the protocol that governs these machine-to-machine interactions. And just as MX thinking applies to MCP server design, it applies equally to A2A: the quality of an agent's experience when interacting with another agent depends on the clarity of that agent's Agent Card, the reliability of its task handling, and the consistency of its outputs.

At Exco Partners, we describe this as designing the social layer of machine experience — the norms, interfaces, and expectations that govern how agents work together. An agent that publishes a vague, incomplete, or unreliable Agent Card offers a poor machine experience. One that is precise, predictable, and well-documented is a first-class citizen in the agent ecosystem.

This has direct business implications. As AI agents become active evaluators and consumers of services — selecting vendors, delegating tasks, and routing work based on capability and quality — the machine experience your systems offer will influence whether those agents choose to work with your organisation or route around it.

Mobile-first design reshaped competitive landscapes a decade ago. The organisations that invested early in machine-first design — including the A2A interfaces their agents expose — will define the leaders of the coming decade. A2A, together with MCP, is the technical foundation on which that advantage is built.