AGTBRIEF
FedAgentKE is a federated knowledge evolution framework that lets AI agent systems share reasoning without sharing raw trajectories, in author run experiments only.
Heterogeneous AI agent systems typically learn in isolation, or share knowledge by forwarding raw reasoning traces that may leak private chain-of-thought. A new arXiv preprint, FedAgentKE, proposes a third path: evolving shared semantic abstractions without ever exchanging raw trajectories.
The paper, posted July 23, 2026, comes from Weihao Li at Northwestern, Jun Bai at McGill and Mila, and Ziyang Song at Ohio University. The mechanism is iterative semantic knowledge distillation, aggregation, and adaptation across distributed agent frameworks (distinct software stacks that wire language models to tools and memory). Each agent distills its experience into a transferable abstraction that is then aggregated and redistributed.
In author-run experiments on the GAIA benchmark, the framework lifted the OWL agent by an average of 19.39 points and SmolAgents by 13.34 points within a single framework, and scaled from 49.0% to 54.4% as cross-framework clients grew from 2 to 8. On SWE-bench Lite, SWE-agent moved from 27.00 to 42.67 and OpenHands from 29.33 to 44.67. Round-wise evolution on GAIA's SmolAgents track moved from 44.24% at round 1 to 74.55% at round 5, per the paper's HTML body.
All numbers are author-reported. The paper is not peer-reviewed, no third-party replication exists, and the framework has only been tested on two benchmarks at small client counts. The authors flag this scope in a limitations section.