stigmergy-autogen

Replace AutoGen's LLM-based speaker selection with deterministic pressure signals. Stop paying for an LLM call every time you need to pick who speaks next. Tasks, priorities, and dependencies drive selection instead.

AutoGen's default selection burns an LLM call per turn to decide who speaks. At 100+ turns that's real money and latency for a routing decision. Stigmergy makes the same decision in sub-millisecond time using pressure signals, and the cost doesn't change at 10,000 turns.

One-line integration with AutoGen v0.7+ SelectorGroupChat.

Install

pip install stigmergy-autogen

Usage

from autogen_agentchat.agents import AssistantAgent
from autogen_agentchat.teams import SelectorGroupChat
from stigmergy_autogen import StigmergySpeakerSelector

# Create AutoGen agents
researcher = AssistantAgent("researcher", model_client=model_client)
analyst = AssistantAgent("analyst", model_client=model_client)
writer = AssistantAgent("writer", model_client=model_client)

# Create stigmergy selector with tasks and dependencies
selector = StigmergySpeakerSelector(wake_threshold=0.4)
selector.register_task("research", agent_name="researcher", priority=0.8)
selector.register_task("analyze", agent_name="analyst", priority=0.6, deps=["research"])
selector.register_task("write", agent_name="writer", priority=0.5, deps=["analyze"])

# Use as SelectorGroupChat selector_func
team = SelectorGroupChat(
    participants=[researcher, analyst, writer],
    model_client=model_client,
    selector_func=selector.select_speaker,
)

result = await team.run(task="Begin research on AI trends")

How It Works

The StigmergySpeakerSelector replaces AutoGen's default speaker selection:

1. Each registered task deposits a pressure signal targeted at its agent
2. On each turn, signals decay and the selector picks the agent with highest pressure
3. Task dependencies are enforced — an agent won't be selected until its deps complete
4. Completion signals propagate pressure to downstream agents

License

MIT