BeeAI now has multiple agents, and a standardized way for them to talk

If two heads are better than one, imagine what you could do with a dozen or more. That’s the inspiration behind IBM Research’s newly redesigned multiagent experimental platform, Bee AI.

BeeAI is aimed at making it easy for developers to run popular open-source AI agents from different frameworks and build specialized agents of their own. No matter where the agents originated, or what code they were written in, they can be easily configured to work alone, or with AI teammates, through the IBM Research-designed agent communication protocol (ACP).

The ACP standardizes how agents talk to each other, removing one of the main barriers to developing multiagent systems. “Right now, agent-to-agent communication is challenged by inconsistent agent interfaces,” said Kate Blair, the director of product incubation at IBM Research who oversees BeeAI. “ACP will act like a universal connector, providing a standardized way for them to exchange information and interact with other systems."

The initial BeeAI experiment was geared toward business users, but BeeAI was revamped earlier this year to focus on developers. Its new goal is to streamline the process of finding, integrating, and orchestrating AI agents, no matter which framework or programming language those agents were built in.

The protocol underlying BeeAI that makes this all possible is built on Anthropic’s model communication protocol (MCP). Introduced last November, the MCP has standardized how agents connect to tools and data to interact with and accomplish tasks in the real world.

IBM’s ACP takes things a step further by introducing the ability to discover and run agents. ACP is currently in the “pre-alpha” stage, as IBM researchers rally the open-source community to build out the protocol to make it easier to discover agents, delegate tasks, and improve agent-to-agent operability. ACP is expected to soon be independent of MCP.

Through BeeAI, users can try out some of the most popular open-source agents on the Web with two commands in the command line interface or a click in the UI. Current offerings include widely used open-source agents like Aider, a pair programmer for editing code, and GPT-Researcher, an agent for gathering and organizing research and citations. There’s even an agent for turning research into structured podcasts optimized for AI-driven text-to-speech to dialogue.

On Friday, Blair showed off BeeAI at the AI Dev 25 conference in San Francisco with Ismael Faro, distinguished engineer at quantum and AI at IBM Research. Hosted by Coursera and DeepLearning.AI founder, Andrew Ng, the conference draws hundreds of developers each year.

“We want to open the discussion, and help move the community toward a universal standard,” said Faro. “We also want to take a feature-driven approach.”

Check out the project on GitHub and read the draft ACP proposal.