Skill∞: an agent improves its own skills and its meta-skill, scoring 60.7% on SkillsBench

A team from the National University of Singapore has proposed a way to close the self-improvement loop where it is easiest to do in practice: at the level of language-agent skills. The Skill∞ system uses a meta-agent. The meta-agent follows a meta-skill to optimize skills for specific tasks, then applies the same procedure to the meta-skill itself. On the 83 tasks of the SkillsBench benchmark, the method reached a 60.7% pass rate. Human-written skills score 56.4%, and the closest automated competitor, SkillEvolver, scores 54.5%. The gap is 4.3 percentage points over humans and 6.2 over SkillEvolver.

Why skills suit recursion

Recursive self-improvement (RSI), in which a system improves not only itself but also the way it improves itself, has long been considered one possible path to AGI. The authors approach it from a practical angle. Language agents increasingly store procedural knowledge as reusable skills: packages of instructions that the agent loads for a given task. If the optimization procedure is also packaged as a skill, it can be improved with the same tools as any other skill.

Existing skill optimizers don't work this way. They follow a procedure designed by a human. Even methods that evolve a meta-skill lock it into a fixed pipeline and retrain it from scratch for each task. Skill∞ removes both constraints: the meta-skill is not tied to a pipeline and is not retrained from scratch.

What the meta-agent does

In Skill∞, the agent itself runs the optimization, not an external script. The meta-skill describes four actions, and self-improvement can rewrite each of them:

• diagnosing execution traces: the agent works out where and why a skill failed on tasks;
• generating candidates: based on the diagnosis, the agent proposes new versions of the skill;
• independent reviews: the agent requests evaluations of the candidates from separate reviewers;
• allocating the search budget: the agent decides for itself where to direct limited resources.

Another distinguishing feature is the decoupling of levels. The meta-skill is separate from task skills. It is refined on several skill-optimization tasks and then reused across different domains without retraining. Skills at both levels remain standard packages and are loaded through the standard skill-discovery mechanism, as in a real deployment. No separate infrastructure is needed to adopt it.

SkillsBench numbers

The comparison used 83 SkillsBench tasks and measured the share of tasks passed:

• Skill∞ — 60.7%;
• human-written skills — 56.4%;
• SkillEvolver — 54.5%.

The main takeaway is not the absolute figure. Automatically optimized skills outperformed hand-written ones, and the procedure that optimized them was itself a product of optimization. The paper's authors are Wangcheng Tao, Zhanglu Yan, Han Wu, and Weng-Fai Wong. The preprint is published as arXiv 2610.35434 and made it into the daily Hugging Face Papers selection.

What teams building agents can take from this

If your stack already has skills in standard packages, two ideas from the paper are worth a look. First, the procedure for improving skills can also be written as a skill rather than hard-coded into the pipeline. It can then be edited with the same tools and from the same traces. Second, independent review of candidates and management of the search budget are better left to the agent itself, since these are steps the meta-skill can redesign for the task.

So far this is a result on a single benchmark of 83 tasks, and transfer to real workloads remains to be tested. But the direction is clear: skill libraries are ceasing to be static folders of instructions and are starting to pull themselves up to a level that used to require manual tuning. Automated optimizers already outperform human-written skills, and built-in self-improvement of the procedure gave them a few more percentage points.