Collabflow improves teamwork among ai agents through learning
CollabFlow: Recursive Self-Improvement of Agent Collaboration
Multiagent SystemsArtificial Intelligence
Summary
Working together well is a challenge for multiple AI agents that need to solve tasks together. The authors designed CollabFlow, a system where a special manager AI builds teams, runs them without change, and learns from how well they do to make better teams next time. The agents only change their minds when given strong evidence from their partners, which reduces mistakes in communication. Over time, the system keeps several good teams active so improvements continue steadily. Tests on many tasks showed that CollabFlow beats existing methods and keeps getting better.
What this means in practice
- •For software engineers: Build AI systems with better teamwork by training a director agent to form and improve collaborative teams automatically.
- •For conversational ai developers: Improve multi-agent dialogue systems by enabling evidence-based communication protocols to reduce error propagation.
Authors
Xiao Huang, Mingda Zhang, Junming Zhang, Qiang Huang, Hanwen Zhang, Yue Dai, Zijia Wang, Xiaoying Tang
Abstract
Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.