A2A is a cross-vendor wire format and discovery layer so agents from different frameworks can hand off tasks without bespoke adapters.
Imagine every shipping company invented its own box shape, and to send a package from FedEx to UPS you had to repack it by hand. A2A is the standard shipping box for agent tasks. An agent built on one vendor's framework can hand a job to an agent on another vendor's framework, and both sides agree on what the box looks like, what is written on the label, and how to confirm delivery. You stop writing a new adapter for every pair of frameworks. It does not change what is inside the box, just how the box travels across vendor boundaries.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Through 2023 and 2024, multi-agent systems lived inside one framework. If you built on LangGraph, your supervisor and workers were all LangGraph nodes. If you adopted CrewAI, your crew shared CrewAI's state object and process model. The reason was not philosophical. It was mechanical: each framework defined its own handoff primitive, its own state representation, its own message schema, and its own discovery convention. Mixing two frameworks meant writing a custom bridge that translated between their internals. The bridges fell out of date whenever either side released a new version.
The A2A protocol is the standardisation effort to fix this at the wire level. Google announced it in April 2025. It is now a Linux Foundation project and is slowly maturing. It pairs with MCP to form a two-layer interop design.
Mental model: A2A is to multi-agent systems what HTTP is to microservices. It does not tell each side how to think or remember; it tells them how to talk to each other.
The N-by-M problem A2A targets
Why every framework had its own handoff
Each multi-agent framework (LangGraph, AutoGen, CrewAI, OpenAI Agents SDK, Mastra, smolagents, MetaGPT) shipped its own primitives because each was designed around its own state representation. LangGraph thinks in typed state dicts with reducers. AutoGen 0.4 thinks in actor messages. CrewAI thinks in role-task pairs. OpenAI Agents SDK thinks in tool calls. The handoff primitive in each fits the local state model. So a cross-framework handoff needs translation.
The combinatorics
With K major frameworks, every interop pair needs an adapter. K of 6 means 15 pairwise adapters. Each adapter must be maintained on both sides whenever either side releases a new version. The bridges fall out of date.
Standardising the wire, not the framework
The insight behind A2A is the same insight behind HTTP, gRPC, and MCP: standardise the wire and let each side keep its internals private. You do not need every agent framework to agree on state representation. You need them to agree on how a task moves between two agents and what comes back. That is a much smaller surface than 'agree on a framework'.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Google announced A2A in 2025 and shipped reference adapters for its Vertex AI agent surface.
- LangSmith Agent Server exposes LangGraph agents as A2A endpoints, so an external agent can treat a whole multi-agent graph as one A2A peer.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does A2A compare to plain REST or gRPC between two agent services?
Both are HTTP-based. The win of A2A is shared semantics for discovery, streaming progress, task lifecycle, and auth, so two strangers can interop without negotiating those layers. Plain REST works fine when both sides are owned by the same team.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Confusing A2A with MCP. MCP is model to tool; A2A is agent to agent. They sit at different layers and solve different N-by-M problems.
60 second bullets to scan on the way to the call.
What A2A standardises (discovery, task envelope, result format)
The N-by-M problem A2A targets across agent frameworks
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.