arXiv 2606.06741's OpenSkill changes the problem setting for "self-evolving Agents" to something cleaner: in real open deployments, often there is no available learning loop — no curated skill library, no successful trajectories, no explicit verification signal; the only thing left to the Agent is the task prompt. OpenSkill lets the Agent acquire "anchored" domain knowledge and verification evidence from documentation, code repositories, and open web pages, then synthesize that knowledge into transferable skills, and use self-built virtual tasks for reinforcement — these tasks are based on anchors, not target answers, completing self-polish without ever touching target supervision.
The paper tests two target Agents on three benchmarks, achieving the highest auto pass rate under the "unsupervised" constraint. Two things are more interesting: the skills learned by OpenSkill can be transferred across models without targeted fine-tuning; its self-built verifiers, although never having seen real results, agree highly with ground-truth judgments.
The point worth saying is: over the past year, self-evolution-related work has almost all been built on the premise of "available supervision first" or "golden trajectory first" — a premise that is hard to assemble in industrial deployment. OpenSkill puts "unsupervised + cross-model transferable" on the table, equivalent to pulling the data threshold for Agent training back to the level of "as long as there are public documents." Combined with the recent trend of frequent self-evolution papers, if this path works, the iteration cycle for enterprise Agents will shorten from "months" to "weeks." But stay cool: the boundary of cross-model transfer on reasoning-heavy tasks, and the robustness of self-built verifiers on long-tail distributions, both still lack systematic data. The algorithm is elegant; the next thing to see is not the paper score, but whether it actually runs in real production environments.