QueenBee Planner slashes token costs by learning when agents should talk

QueenBee Planner, developed by researchers at an unnamed institution, tackles a core inefficiency in LLM multi-agent systems: how agents are connected. Rather than hand-designing or randomly assigning who talks to whom, QueenBee treats the communication topology as a retrievable, self-improving skill. A frozen pool of worker agents and a task adapter generate execution traces; an outer LLM planner learns to produce temporal directed acyclic graphs (DAGs) specifying information flow per round. The system distills execution traces into evidence-backed design rules with three actions: Preserve, Modify, or Avoid. To prevent self-evolution from overfitting to lucky runs, it uses held-out acceptance gates, variance-aware credit, motif-level attribution, and structural deduplication.

In evaluations on count-frequency aggregation and Silo-Bench-style distributed coordination tasks, QueenBee's self-evolved graphs outperformed all fixed topologies and cold (non-evolved) generation. On the CF fulltest set, RMSE dropped from 12.53 (best fixed topology) to 7.87, while simultaneously reducing the number of messages, model calls, and token cost. The same direction of improvement held for Silo-style tasks. These results demonstrate that multi-agent systems can learn reusable architectural design knowledge rather than merely memorizing task answers — a step toward more efficient and adaptive AI collaboration.