Bee Agent Framework

Bee Agent Framework is an open-source TypeScript framework developed by IBM Research for building production-grade AI agents. It provides a structured approach to agent development with first-class TypeScript support, making it a compelling choice for teams building agents in the JavaScript/TypeScript ecosystem who want enterprise-quality tooling backed by IBM's AI research.

The framework centers around a ReAct-style agent loop where agents reason about tasks, take actions through tools, and observe results iteratively. Bee provides a rich set of built-in tools including web search, code execution (via Python and JavaScript sandboxes), Wikipedia access, weather, and more. Custom tools are defined with TypeScript interfaces and Zod schemas, providing full type safety and automatic schema generation for LLM function calling.

Bee's memory system is designed for production use with multiple memory types: sliding window memory for recent context, token-based memory for fitting within model limits, and summarization memory for compressing long conversations. These can be combined for sophisticated memory strategies that balance context quality with token efficiency.

Observability is built in through the Emitter system, which produces structured events for every step of agent execution — LLM calls, tool invocations, memory updates, and errors. These events integrate with standard observability tools for monitoring production deployments. The framework also provides serialization for saving and restoring agent state.

Bee supports multiple LLM providers including IBM watsonx.ai, OpenAI, Ollama, and Groq through its adapter system. The framework integrates with LangChain tools, allowing teams to leverage LangChain's tool ecosystem while using Bee's agent loop and memory management. IBM actively maintains the project with regular releases and production use in IBM's own AI products.