A new arXiv paper separates two jobs that are often blurred in agent systems: following a skill and improving the skill itself.
The authors note that expert-written natural-language skills can help tool-using agents, but agent-authored skills perform 8 to 11 points worse than using no skill. Their framework, WER, short for Write, Execute, and Refine, trains a skill optimizer while keeping the executor model frozen.
The optimizer proposes skills, the frozen agent tries them repeatedly, and a programmatic verifier scores the outcomes. Mixed success and failure cases become training states for later refinement, so the optimizer learns from execution evidence rather than merely rewriting instructions at inference time.
The work matters because many agent products depend on procedural guidance, playbooks, and tool recipes. If agents can reliably improve those instructions from test results, developers could get more durable gains than one-off prompt repair loops. The paper remains an experimental framework, not a guarantee that self-written skills are safe or generally useful.