BeeAI Framework

The main agent type is RequirementAgent: a tool-calling loop whose choices are narrowed by requirements. A ConditionalRequirement can force a tool at a given step, allow it only after another tool, forbid consecutive calls, or set minimum and maximum invocation counts; middleware such as GlobalTrajectoryMiddleware observes every step through the event emitter.

Multi-agent setups come in two forms. With HandoffTool, a lead agent treats specialist agents as tools, calls them, and uses their answers in its own reply (the README's multi-agent example). With workflows, a Workflow holds a typed state model and named steps, each returning the next step or NEXT, SELF or END, and workflows can nest; AgentWorkflow adds agents as sequential steps so a final agent can combine earlier results.

Each agent carries a memory object: UnconstrainedMemory, SlidingMemory, TokenMemory or SummarizeMemory. Longer-lived state is handled by the Serializer, and servers such as A2AServer keep per-session memory through a memory manager (an LRU manager in the example).

In Python, AskPermissionRequirement marks tools that need approval before they run; by default the approval is a simple prompt in the terminal, and a custom handler can route the decision to an external system or UI. The docs state this requirement is not yet implemented in TypeScript. No other interrupt or edit-state mechanism is described on the pages read.

pip install beeai-framework

import asyncio

from beeai_framework.agents.requirement import RequirementAgent
from beeai_framework.backend import ChatModel
from beeai_framework.tools.handoff import HandoffTool
from beeai_framework.tools.weather import OpenMeteoTool


async def main() -> None:
    llm = ChatModel.from_name("ollama:granite4.1:8b")  # local model served by Ollama
    weather = RequirementAgent(llm=llm, tools=[OpenMeteoTool()], role="Weather Specialist",
                               instructions="Give short forecasts for a named city.")
    lead = RequirementAgent(name="Lead", llm=llm, tools=[
        HandoffTool(weather, name="WeatherLookup", description="Ask the weather specialist.")])
    result = await lead.run("Should I pack an umbrella for Rome this weekend?")
    print(result.last_message.text)


asyncio.run(main())

• The Python package requires Python >=3.11 and <3.14.
• Protocol integrations are extras: pip install 'beeai-framework[a2a]', [mcp], [agentstack] or [acp-zed].
• The README example needs Ollama running with granite4.1:8b pulled; hosted models use provider:model names such as openai:gpt-5-mini plus the provider's API key.
• The docs quickstart uses the starter repositories (uv sync for Python, npm ci for TypeScript) rather than a bare install.
• AskPermissionRequirement is not implemented in TypeScript, and an A2A server hosts only one agent.

Yes to both. MCPTool and MCPServer cover consuming and exposing MCP tools; A2AAgent and A2AServer cover calling and publishing A2A agents. AG-UI support was not found.

Yes. It is Apache-2.0 licensed and no paid tier of the framework is documented. You pay only for any hosted model you choose; local models run through Ollama.

Python (3.11 to 3.13) and TypeScript. Some features, such as AskPermissionRequirement, exist only in Python so far.

The README's legal notice says IBM has no obligation to support the code and will not maintain it going forward; development continues through BeeAI contributors under the Linux Foundation AI & Data program, with releases as recent as September 2026.

Either a lead agent calls specialist agents through HandoffTool, or a Workflow/AgentWorkflow runs agents as steps over a shared typed state.