ACP and MCP are tailored for their respective roles. MCP uses JSON-RPC as its communication protocol, while ACP adopts a RESTful architecture implemented over HTTP, supporting both synchronous and asynchronous agent interactions.
ACP allows developers to interact with agents directly with tools like curl, Postman, or a simple web browser, without a dedicated software developer kit (though an SDK for Python and TypeScript are provided). ACP also allows agents to carry their own metadata so they can be found even in secure or air-gapped setups.
ACP opens new architectures beyond the “manager” pattern where a boss agent calls other agents. With ACP, agents can interact as peers, instead of through an intermediary. This matters if agents within an organization, or from outside, need to handle business on their own.
“Either one should be able to initiate a conversation or delegate a task,” says Blair. “Within an organization, you might have an agent triaging customer queries that should be able to send their customer to the right service agent which can then close out the ticket.”
The BeeAI team showed off an early version of ACP in March, at the AI Dev 25 conference in San Francisco hosted by tech entrepreneur and DeepLearning.AI founder Andrew Ng. A few weeks later, Google unveiled its own agent-to-agent protocol, A2A.
Is there room for two agent protocols? “Absolutely,” says Blair, “at least for now. It’s early days and I expect there will be a lot of iterations as protocols are put to the test under real-world use.”
As a community-led project with an open governance structure, ACP allows developers to adopt and improve on the standard without worrying about vendor lock-in. “We have monthly open community calls and an active GitHub discussion section,” says Blair. “We make sure the project always has tasks for community members to pick up and pitch in.”
And ACP is already part of the AI curriculum. Starting June 25, the Linux Foundation and BeeAI will offer a short course on ACP through Ng’s education platform, DeepLearning.AI. The ACP course will complement an existing course on MCP and be taught by IBM AI engineers Sandi Besen and Nicholas Renotte. Mark your calendar.