Definition: Agent Communication Protocol (ACP)
Agent Communication Protocol (ACP) is an open, REST-based standard, built by IBM Research for the BeeAI platform, that lets autonomous AI agents on different frameworks discover each other and exchange tasks over one HTTP interface.
Core characteristics of Agent Communication Protocol (ACP)
ACP standardizes agent-to-agent messaging on plain web conventions instead of a bespoke format, so any system that calls a REST endpoint can take part.
- Unified RESTful interface shared by every agent, regardless of framework
- Agent Detail manifest for online or offline discovery of capabilities
- Synchronous calls plus Server-Sent Events streaming for long tasks
- Stateless-by-default servers built to scale behind load balancers
Agent Communication Protocol (ACP) vs. Model Context Protocol (MCP) and Agent-to-Agent Protocol (A2A)
ACP, MCP, and A2A appeared within months of each other in 2024 and 2025 but solve different layers of one problem. MCP connects a single agent to its tools and data, acting as a client. ACP and A2A both connect one agent to another. The difference between ACP and A2A is now mostly historical: ACP came from IBM’s BeeAI project on REST, A2A from Google on JSON-RPC. In August 2025 their designs merged under Linux Foundation governance, so new work targets A2A while ACP remains the base many existing BeeAI agents run on.
Importance of Agent Communication Protocol (ACP) in enterprise AI
ACP mattered as one of the first attempts to standardize agent-to-agent communication on plain REST, easy for existing API teams to adopt. An arXiv survey in 2025 named ACP alongside MCP, A2A, and ANP as one of four protocols enterprises actively evaluated (source: arXiv, 2025). Bitkom’s February 2026 guide on multi-agent ERP systems treats this interoperability as a precondition for cooperating multi-agent systems in Germany (source: Bitkom, 2026).
Methods and procedures for Agent Communication Protocol (ACP)
Working with ACP means understanding its original design even as deployments migrate to its successor.
Agent manifest publication and discovery
Every ACP agent publishes a manifest describing what it does and which inputs and outputs it expects.
- Define the agent’s input and output schema
- Publish the manifest online, or package it offline for idle agents
- Version the manifest so callers catch breaking changes early
Synchronous and streaming execution
Clients can call an agent once for a short task or stream a long one. The stateless-by-default server lets one instance serve many callers without tracking per-conversation state, simplifying scaling behind load balancers.
Migration path toward A2A
ACP’s stateful and asynchronous task concepts were folded into A2A rather than discarded. Its reference repository was archived in August 2025, so new integrations target A2A while legacy ACP endpoints can run behind a translation layer during the transition.
Important KPIs for Agent Communication Protocol (ACP)
Measuring an ACP-derived deployment focuses on reliability, adoption, and discovery quality.
Interoperability reliability metrics
- Agent call success rate: above 99%
- Discovery resolution time: under 1 second
- Stream completion rate: above 98%
- Cross-framework compatibility: share needing no custom adapter
Strategic adoption metrics
Tracking how many agents expose a manifest versus still needing bespoke integration shows progress away from point-to-point links. Analysis citing Gartner’s 2026 research projects over half of enterprise agents will run on standardized frameworks by 2027 (source: Gartner, cited 2026).
Quality metrics
Quality tracking should cover whether a caller got the expected output format on the first attempt, since manifest mismatches between frameworks are a common silent-failure source.
Risk factors and controls for Agent Communication Protocol (ACP)
ACP’s status as a protocol in transition brings risks a stably maintained standard would not.
Building on a protocol in transition
Agents built on the original spec run on an archived reference implementation, a growing gap for teams without a migration plan.
- Audit which agents still depend on ACP-specific endpoints
- Build a translation layer so ACP clients can call A2A agents
- Set a target date to retire direct ACP dependencies
Unverified identity across frameworks
Because ACP, like A2A, lets agents from different vendors call each other, one that skips identity checks exposes itself to a compromised counterpart. Deployments should authenticate every call, log the caller’s identity, and restrict permitted task types.
Stateless and stateful design mismatch
ACP assumes each call carries its own context, but some workflows, such as a multi-turn negotiation, are inherently stateful. Bolting session tracking onto a stateless layer without clear ownership creates bugs where agents disagree on the task’s current step.
Practical example
A 140-person industrial automation supplier in Bavaria builds machine-vision modules its customers integrate into their production lines. Before an agent-to-agent standard, its support agent and each customer’s maintenance agent only exchanged information by forwarded email, since the systems ran on different frameworks. After the supplier published an ACP-style manifest for its diagnostic and parts-lookup capabilities, customer agents on other frameworks could call it directly over the same interface.
- The maintenance agent submits a fault code and gets a diagnosis in seconds
- Parts and lead-time queries route to the ERP-connected agent, not a mailbox
- Unresolved escalations go to a human engineer with the full exchange attached
- The supplier is mapping the same integration onto A2A for new customers
Current developments and effects
ACP’s 2025-2026 story is less about growing alone and more about its ideas being absorbed elsewhere.
Convergence with A2A under Linux Foundation governance
In August 2025, IBM and Google announced ACP’s design, including stateful and asynchronous task handling, would merge into A2A rather than continue separately.
- IBM’s Kate Blair joined the A2A technical steering committee after the merger
- A2A became a hosted Agentic AI Foundation project in August 2026
- Both protocols stay documented, but new work targets A2A
BeeAI and the open agent ecosystem
ACP still powers BeeAI, IBM’s open-source platform for discovering and composing agents, even as governance moved elsewhere. Existing BeeAI-based agent orchestration need not rebuild at once, but new links increasingly use A2A-compatible interfaces.
Disambiguation from Zed’s Agent Client Protocol
The same acronym also names Zed’s unrelated Agent Client Protocol for coding agents and editors. The two share no lineage, so confirming which ACP a document means is worth the extra sentence.
Conclusion
Agent Communication Protocol mattered because it proved, alongside MCP and A2A, that agent-to-agent communication could run on conventional web technology rather than a proprietary format per vendor. Its legacy lives on inside A2A, which absorbed ACP’s design instead of starting over. For Mittelstand enterprises weighing workflow automation with customers’ or suppliers’ own agents, the sensible move now is building new integrations against A2A directly. Enterprises already running ACP-based agents have a defined, Linux Foundation-governed path to migrate, not a dead end.
Frequently Asked Questions
What is the difference between ACP and MCP?
MCP connects a single AI agent to the tools and data it needs, as a client to specialized servers. ACP, like A2A, connects one agent to another so two systems can exchange tasks directly.
Is ACP still used, or has it been replaced by A2A?
ACP’s original implementation was archived in August 2025 after merging into A2A under Linux Foundation governance. Agents built on the original spec, including BeeAI ones, keep running, but new work should target A2A directly.
Does adopting an agent communication protocol make sense for a company with around 150 employees?
Yes, if its agents must exchange tasks with agents run by customers, suppliers, or another department on a different framework. Once more than one or two counterparts are involved, a shared protocol beats custom integration per partner.
What does implementing an agent-to-agent protocol cost, and do we need our own IT resources?
Connecting to a partner that already publishes a compliant manifest is mostly configuration, not development. Building manifest support into an internal agent typically takes one to three weeks, an investment that carries forward to every later partner.
How does this fit with DSGVO and the EU AI Act?
Agent-to-agent protocols are communication layers, not compliance frameworks, so tasks carrying personal data across company borders still need a DSGVO assessment. EU AI Act Article 14 requires traceable logs for high-risk decisions, so shared logging agreements with partners matter alongside the protocol’s own authentication.
Where can I still find documentation for ACP if we are maintaining an older BeeAI-based agent?
IBM Research’s project page and the archived reference repository remain available as read-only references. For anything new, IBM points to A2A, which incorporated ACP’s key decisions, as the maintained standard.
