Researchers have found a way for AI agents to get smarter by comparing notes, without ever showing each other their homework.
The framework, called Federation over Text (FoT), has multiple LLM agents work on separate tasks, then send their reasoning traces, not the underlying problems, to a central server. That server distills the traces into a shared library of insights that other agents, current or future, can pull from to reason better. Unlike federated learning, which averages model gradients during training, FoT operates purely on text after the fact, with no gradient optimization or supervision signal involved. Across tests on everyday tasks, cross-domain collaboration, and research insight discovery, the approach lifted performance by an average of 11.9 percentage points while cutting the number of tokens agents needed to finish tasks by 5.5%.
Most agent systems today are specialists that never talk to each other: a coding agent and a scheduling agent don't trade lessons learned. FoT's pitch is a lightweight way to pool know-how across domains without the privacy and bandwidth costs of sharing raw data or retraining a shared model, which matters as companies deploy more narrow, task-specific agents that could otherwise never benefit from each other's experience.
The 11.9-point gain is an average across the paper's own benchmark suite, so treat it as a promising lab result rather than a settled number until independent teams test it on harder, messier tasks.