ACP (Agent Communication Protocol)

What it is

ACP, short for Agent Communication Protocol, is an open protocol for agent interoperability. Its job is to connect AI agents with each other, with applications, and with humans ^[1]. In practice that means an agent built by one team can be called by another agent, invoked from an application, or driven by a person, without either side having to know how the other was built.

The problem it addresses is the same one that any interface standard addresses: without a shared convention, every pairing of agents needs its own custom glue, and a system with many agents ends up with many one-off connections. A single agreed protocol turns that into one interface each participant has to implement.

How it works

ACP is built around a standardised RESTful API ^[1]. A REST API is the ordinary request-and-response style of web interface, so callers reach an agent over conventional HTTP calls rather than through a bespoke transport that each vendor would have to define.

The protocol supports both synchronous and asynchronous communication ^[1]. A caller can send a request and wait for the answer, the way a function call returns a value, or it can hand off a longer-running job and collect the result later without holding the connection open.

It also supports both stateful and stateless operation patterns ^[1]. In a stateful pattern the agent keeps context across successive calls, so a conversation can build up over several turns. In a stateless pattern each call carries everything it needs and leaves nothing behind, which makes calls easier to retry, spread across servers, or scale independently.

Agents can be discovered either online or offline ^[1]. Online discovery means the agent can be found while the system is running; offline discovery means its details are known in advance, so the caller goes straight to it rather than looking it up at runtime.

ACP is deliberately agnostic to internal implementations ^[1]. It does not prescribe how an agent is put together, and it works with agents built with LangChain, CrewAI or entirely custom code ^[1]. Anything that can speak the API can take part.

To make that easier, the project publishes Python and TypeScript SDKs, a REST API and an OpenAPI specification ^[1]. The SDKs give developers a language-native way to expose or call an agent; the OpenAPI specification describes the HTTP surface in a machine-readable form, so other tools can generate clients or documentation from it.

Where it came from

ACP was developed as an open standard under the Linux Foundation, alongside BeeAI ^[1]. The protocol and the platform it serves were therefore developed together, rather than the protocol being retrofitted around an existing product.

Where it is used

ACP underpins the BeeAI platform, which lets teams test, deploy and share ACP agents ^[1]. That covers the practical loop of the work: trying an agent out, putting it into service, and making it available to others.

Current status

The ACP documentation now states that ACP is part of A2A under the Linux Foundation, and links a migration guide ^[1]. Readers who arrive at the ACP documentation are pointed toward A2A as the home of the standard, with a documented route for moving existing work across.