A new arXiv paper proposes a way for AI agent teams to retrain themselves after every task, not just within one.
Researchers describe CollabFlow, a system for multi-agent LLM collaboration. A trainable component called the Collab-Director assembles teams of agents; a separate, frozen executor actually runs those teams on tasks. After each round, the results retrain the director, so team composition improves over time instead of staying fixed by whoever set it up. Within a team, an agent only switches to a teammate's answer when that teammate's evidence is clearly stronger, which is meant to stop one agent's mistake from spreading unchecked through the group.
Most multi-agent AI setups today have a human operator decide who talks to whom and when, a slow, manual, one-time design choice. CollabFlow argues that collaboration itself can be learned and continuously rewritten based on outcomes, and its training objective, called Collaborative Trajectory Balance, deliberately keeps several good team configurations alive rather than collapsing onto one favorite, which the authors say prior reward-maximizing approaches tended to do.
The claim, outperforming baselines across twelve datasets and continuing to improve round over round, is the kind of result that is easy to produce on curated benchmarks and harder to replicate once real-world messiness and cost constraints show up. For now the code sits behind an anonymized repository link, standard for anonymous peer review, not a public release.