Maosen Zhou is a co-author of research on tool-using agents, including WEFT, ContextWeave and the Nex-N1 report. These projects study how training environments, execution systems and memory affect an agent’s ability to complete longer tasks.[1]
Start with WEFT →Contributions
Improve the whole training interaction
WEFT treats the environment, task, agent harness and evaluator as a coupled system. Execution traces and state evidence help identify which component caused a failure; revised components are then tested through new rollouts. Its training approach preserves verified progress and assigns learning signals to individual interaction turns. Zhou is one of the paper’s co-authors, rather than the sole creator of this system.[1]
Measure whether experience remains useful
Zhou also co-authored ContextWeave, which evaluates agent memory through continuing workflows. Its tests distinguish retrieving relevant experience from successfully applying it, including whether misleading recall damages the final result. This connects memory design with the same execution-level outcomes that matter in agent training.[2]
Tools, writing & talks
Writing, research & talks
Start here
WEFTarxiv.orgA whole-system approach to constructing and improving tool-use training interactions.
Begin with the four components of an agentic interaction system, then follow one execution-driven revision cycle. Familiarity with reinforcement learning helps with the training details. Compare the reported results within the paper’s matched-model settings.
Co-author[1]- ContextWeavearxiv.org
Workflow-based evaluation of agent memory and the effects of misleading recall.
Co-author[2] - Nex-N1 technical reportarxiv.org
A collaborative technical report on environment-based agent training.
Co-author[3]