BeeAI Framework
also known as Bee Agent Framework
Type: full-code · Vendor: Linux Foundation AI & Data (originally IBM) · Language: Python, TypeScript · License: Apache-2.0 · Status: active · Status in practice: emerging
Provide a multi-language (Python + TypeScript) framework for production-ready multi-agent systems with a RequirementAgent that enforces declared rules across LLMs, multi-agent workflows, event-driven observability, agent state serialisation, MCP/A2A protocol support, and a unified backend over multiple LLM providers.
Description. BeeAI (formerly the IBM Bee Agent Framework) is the Apache-2.0 multi-agent framework now governed under the Linux Foundation AI & Data program. Tagline: 'Build production-ready multi-agent systems in Python or TypeScript.' Its headline abstraction is the RequirementAgent — a rule-driven agent that produces 'predictable, controlled behavior across different LLMs by setting rules the agent must follow'. The framework ships unified Backend connectors, built-in tools (web search, weather, code execution) plus custom tool support, RAG with vector stores, conversation Memory, multi-agent Workflows, event-based observability, agent serialisation for resumption, and protocol support for MCP and A2A.
Agent loop shape. Rule-driven tool-calling loop. A RequirementAgent runs an LLM-backed loop in which each step is checked against declared requirements (rules) before tools are called; the unified Backend abstracts the underlying provider so the same rules apply across LLMs. Multi-agent topologies are built on top via Workflows that orchestrate several agents. Event emitters fire on every step for logging, tracing, and error handling. Agent state is serialised via Save/load for cross-session persistence; MCP and A2A protocols expose and consume external agents and tools.
Primary use cases
- rule-bound agents that behave consistently across LLM providers
- multi-agent workflows with complex execution flows
- agents that persist state across sessions (serialisation)
- MCP- and A2A-integrated agents
- event-driven observability of agent behaviour
Key concepts
- RequirementAgent (docs) — Rule-driven agent producing predictable behaviour across LLMs.
- Workflows → orchestrator-workers — Orchestrate multi-agent systems with complex execution flows.
- Backend — Unified interface connecting to various LLM providers.
- Memory — Conversation history management with multiple strategies.
- Serialization → agent-resumption — Save and load agent state for persistence across sessions.
- Observability events — Event-based monitoring, logging, and error handling.
- MCP + A2A → mcp — Host agents in servers with support for multiple protocols including MCP and A2A.
Patterns this full-code implements —
- ★★Agent Resumption
Serialization saves and loads agent state across sessions.
- ★★Agentic RAG
Retrieval-augmented generation systems with vector stores are a documented capability.
- ★★Code Execution
Code execution is a built-in tool alongside web search and weather.
- ★★Event-Driven Agent
Event-based monitoring, logging, and error handling are the documented observability model.
- ★★Model Context Protocol
Hosting agents in servers with support for multiple protocols such as A2A and MCP.
- ★★Orchestrator-Workers
Workflows orchestrate multi-agent systems with complex execution flows.
- ★★ReAct
Agents 'capable of reasoning, acting, and adapting' (the ReAct pattern) sit alongside the rule-bound RequirementAgent.
- ★★Tool Use
Built-in plus custom tools are a top-level capability.
- ★★Structured Output
RequirementAgent rules constrain outputs and expected_output accepts Pydantic models or JSON Schema; output_structured is exposed on responses.
- ★★Step Budget
RequirementAgent and base agents expose max_iterations and total_max_retries for explicit step and retry caps.
- ★Handoff
HandoffTool delegates between specialized agents in multi-agent setups (e.g. KnowledgeLookup / Weather Agent handoffs).
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified