AI agents can't teach themselves new tricks – only people can
The Register
February 19, 2026
AI-Generated Deep Dive Summary
AI agents are highly dependent on human-curated skills to perform effectively, according to a recent study. While these agents can handle tasks beyond their training data by accessing new "skills"—such as instructions or reference materials—they struggle when asked to develop these abilities independently. The study, conducted by a team of researchers including Xiangyi Li from BenchFlow and others from leading tech companies and universities, found that AI agents using human-curated skills outperformed those without skills by 16.2% on average across various tasks. However, relying solely on self-generated skills yielded disappointing results, highlighting the limitations of AI's autonomous learning capabilities.
The researchers developed a benchmark called SkillsBench to evaluate how well agent skills enhance performance during inference. They tested seven agent-model setups across 84 diverse tasks, generating over 7,300 trajectories—each representing an agent’s attempt to solve a task under specific skill conditions. Three scenarios were analyzed: no skills, curated skills designed by humans, and self-generated skills developed by the AI itself. The results showed significant variance in performance, with agents using curated skills consistently outperforming those without skills, particularly in specialized domains like healthcare and manufacturing.
For example, in a flood-risk analysis task, agents without any skills achieved only a 2.9% pass rate due to their inability to apply appropriate statistical methods. However, when equipped with a curated skill that provided specific guidance on using the Pearson type III probability distribution and USGS methodology, the pass rate jumped to 80%. This demonstrates how human-designed skills can bridge gaps in AI's inherent knowledge, significantly improving task success rates.
The findings underscore the importance of human involvement in shaping AI capabilities. While AI agents are powerful tools, their effectiveness heavily depends on the quality and specificity of curated skills provided by humans. This matters
Verticals
tech
Originally published on The Register on 2/19/2026