AutoGen integration example

AgentWeave can be used as a pre-team routing layer for AutoGen AgentChat.

The integration boundary is:

user task
    ↓
AgentWeave requirement analysis + ranking + team selection
    ↓
selected AutoGen participants
    ↓
normal AutoGen team execution

This pattern keeps AutoGen responsible for agent execution and conversation while AgentWeave decides which specialists should enter the team for a particular task.

What the example demonstrates

examples/autogen_agentweave.py creates four AutoGen specialists:

• backend specialist;
• database specialist;
• research specialist;
• MCP specialist.

For each task, AgentWeave:

1. infers the task requirements;
2. ranks the available specialist profiles;
3. selects a bounded team;
4. produces a structured selection explanation;
5. passes only the selected specialists into an AutoGen RoundRobinGroupChat.

The example uses AutoGen’s current BaseChatAgent and RoundRobinGroupChat APIs. The specialist agents are deliberately deterministic, so the example can be run without an API key or model service.

Run it

From the repository root:

python -m venv .venv
source .venv/bin/activate
pip install -e .
pip install -U autogen-agentchat
python examples/autogen_agentweave.py

AutoGen AgentChat currently requires Python 3.10 or later; AgentWeave itself requires Python 3.11 or later, so Python 3.11+ satisfies both.

Production pattern

In a real AutoGen application, replace the deterministic SpecialistAgent objects with your normal AssistantAgent, custom AgentChat agents, or teams.

The routing step remains the same:

selected_names, explanation = route_autogen_participants(task)
selected_agents = [all_autogen_agents[name] for name in selected_names]
team = RoundRobinGroupChat(selected_agents, max_turns=...)
result = await team.run(task=task)

This means an application can maintain a large catalog of AutoGen specialists while constructing a smaller task-specific team before conversation begins.

Why keep the selection explanation?

The returned AgentWeave explanation records which candidates were considered, which were selected, capability matches/misses, scores, and other selection information. That makes it easier to distinguish orchestration decisions from downstream AutoGen behavior.

For tool-level routing provenance, see Issue #22.

Evidence and ecosystem boundary

This is an integration example, not a benchmark result. It does not modify AutoGen, does not imply Microsoft or AutoGen endorsement, and does not alter any frozen AgentWeave research evidence.