BeeAI: IBM's open-source bet on the future of multiagent systems

AI agents are all the rage these days. These programs, which can perform tasks on behalf of their users, are considered the future of AI. And they’re becoming a central part of AI strategy for major tech companies, including Oracle, Microsoft, Salesforce and, of course, IBM, not to mention AI giants like OpenAI and Perplexity.

According to Salesforce CEO Marc Benioff, whose enthusiasm was widely reported by the press last December, using AI agents to solve problems and make decisions could be a trillion-dollar opportunity.

Enter BeeAI. Last year, IBM Research released the Bee Agent Framework, a fully open-source and no-code platform to get started with agents. The AI agents, called "bees," connect to an LLM and can access tools to respond to user queries and perform tasks. The agents can also reflect on what they are doing and come up with new approaches. Now, the team is working and releasing enhancements to BeeAI, adding a Python framework, multiagents and improvements to developer experience.

The BeeAI updates push the idea further with multiagent extension. The overarching vision, according to Michael (Max) Maximilien, a Distinguished Engineer at IBM, is to “make simple things simple and complex things possible.”

“The way you think of BeeAI is [that] it's expanding this model where you don't necessarily need just one agent to answer a question. You may need multiple agents.”

The agentic approach tries to mimic how humans interact and get work done in real life.

Typically, a high-performing team consists of people who are highly focused on specific tasks, but who can also manage parallel dependencies being worked on by other team members. To be effective, workers must perform some work concurrently, while other tasks require sequential execution.

“BeeAI allows not only multiple agents, but also agents from different implementations,” Maximilien explains. “They don't all have to be the same type of agent. The idea is that they can collaborate to either answer queries or execute workflows for users.”

There’s also an element of reuse with multiagency, Maximilien says—that is, multiple specialized agents work together rather than one agent trying to do it all. One key aspect: BeeAI is open source, and implemented in TypeScript and Python. The team behind Agent Bee Framework and BeeAI wants to incorporate user feedback. “We want the platform to be out-of-the-box useful, so we want input of the people using it,” he says.

“We have a set of clear opinions on how multiagency should be done, but we also believe it’s best done in the open,” says Maximilien. “It’s similar to how programming evolved from procedural to object-oriented languages. These new paradigms succeeded because many languages, like Java, embraced openness, fostering creativity and ease of use.”

Many organizations are already exploring how to use multiagent frameworks to achieve greater scale and performance—particularly when it comes to executing more complex or domain-specific tasks.

“Most companies need to solve specific problems. To solve those problems with AI, they need to create an agentic solution and codify the workflow using the LLMs and the tools,” Maximilien says. “They need to solve their everyday workflows, empower their users and enhance their operations.”

Organizations see AI not just as a way to automate workflows, but also to empower employees, he believes. That’s the real value.

“We see this as something that will continue to grow with the multiagent additions because individual users will find more value in solving more complex problems,” he says, inviting developers to use BeeAI and integrate it into their tools.