Agent Communication Protocol (ACP)

The Agent Communication Protocol (ACP) was an open protocol from IBM Research for communication between AI agents, applications and people, and it is no longer developed on its own because it merged into A2A.

History. IBM Research launched ACP in March 2025 to power its BeeAI Platform. Later that month IBM donated the BeeAI project, ACP included, to the Linux Foundation. On 25 August 2025 the ACP team announced that ACP was merging with A2A under the Linux Foundation, that it would wind down active development, and that it would contribute its technology to A2A instead. A joint IBM Research and Google post on the LF AI & Data blog confirmed the merger on 29 August 2025. The BeeAI Platform moved to A2A. The ACP repository is archived, and its last release was v1.0.3 on 21 August 2025.

How it worked. ACP was a REST protocol described by an OpenAPI document. Its main concepts were:

• Agent manifest: an agent’s name, description and metadata, used for discovery.
• Run: one execution of an agent with specific inputs, returned synchronously or as a stream.
• Message and MessagePart: ordered, multimodal content, each part tagged with a content type.
• Await: a run pauses to ask the client for more input, then resumes.
• Session: state and history kept across runs.

A client listed agents with GET /agents and started a run with POST /runs. The request body from the project’s quickstart:

{"agent_name": "echo", "input": [{"role": "user", "parts": [{"content": "Howdy!", "content_type": "text/plain"}]}]}

Counterparts in A2A. Several ACP ideas have close A2A equivalents. The manifest resembles an Agent Card, a run resembles a task, Await resembles the input-required state, and a session resembles a contextId. A2A also defines signed Agent Cards and three standard protocol bindings.

Name clash. OpenAI and Stripe’s Agentic Commerce Protocol also abbreviates to ACP. The two are unrelated: IBM’s ACP carried messages between agents, while the commerce protocol runs checkout.