SkillHone stores an agent’s skill-edit history so later runs don’t relearn the same bugs

机器之心 · wechat · 2026-07-24

SkillHone keeps an agent’s skill history so future runs stop relearning the same mistakes

Machine Heart covers SkillHone, a Tencent paper/framework for continual skill evolution in agent workflows.

The core problem: a skill may work this week and fail next week because APIs, rate limits, and web layouts change, while the next agent only receives the final file and loses the reasoning trail.
SkillHone stores a persistent decision history for each optimization round: diagnosis, candidate edits, evaluation evidence, and final acceptance/rejection.
It also expands optimization from a single SKILL.md to the entire skill repository, including scripts, references, and templates.
The system separates optimizer and evaluator roles, supports multiple agent runtimes, and can be wired into existing toolchains.
On GAIA and WebWalkerQA-EN, the paper reports large gains over Hermes-SE and a deep-research baseline; removing decision history causes a big drop in accuracy.
A key result is that the same skill package can transfer to another model without re-optimizing from scratch.

More from coding & agent