Search papers, labs, and topics across Lattice.
SkillHone introduces a novel framework for the continual evolution of agent skills by maintaining a persistent decision history, which allows agents to build upon prior revisions and evaluations. This approach contrasts with existing methods that discard decision histories after bounded runs, limiting the ability to refine skills over time. The results demonstrate that SkillHone significantly outperforms both a commercially backed deep-research agent and previous skill-evolution methods, achieving notable improvements in accuracy across various benchmarks and internal scenarios.
SkillHone enables agents to evolve their skills continuously by leveraging a structured decision history, leading to a remarkable 15.8-point performance boost over existing deep-research agents.
Agent skills extend language-model agents with task-specific procedures, scripts, and references, but the tasks and environments they target continually change. Existing methods improve skills in bounded runs and retain only the final artifact, discarding the decision history that later agents need to interpret prior revisions, evaluations, and rejected alternatives. We introduce SkillHone, a harness for continual agent skill evolution grounded in persistent decision history. SkillHone pairs skill revisions with evaluation-side evidence that supplies practice feedback, recording structured histories of diagnoses, revisions, evidence, and outcomes. Role-separated subagents run candidate skills on practice probes with redacted reporting and propose revisions informed by prior decisions, enabling cross-session refinement without rediscovering past rationale. On deep-research benchmarks, SkillHone runs without a pre-integrated search stack and outperforms the commercially backed deep-research agent by 15.8 points on GAIA and 3.2 points on WebWalkerQA-EN, while also exceeding prior skill-evolution methods. We further deploy SkillHone on internal tool-mediated analysis scenarios, where it improves accuracy by an average of 18.8 points across seven settings.
