- A Princeton and UC San Diego study finds “skills” help AI agents mainly via structured workflows, not added knowledge.
- As the skill library grows, agents have a harder time selecting the right skill for a task.
- The finding tempers the industry push to give agents ever-larger libraries of reusable skills.
- The Decoder’s Maximilian Schreiner reported the study on August 22, 2026.
What Happened
A study from researchers at Princeton University and UC San Diego finds that so-called skills make AI agents better mainly through structured workflows, not through added knowledge, as reported by The Decoder‘s Maximilian Schreiner on August 22, 2026. But the same work finds a failure mode: as the skill library grows, agents have a harder and harder time finding the right set of skills for a given task.
Why It Matters
“Skills” — reusable, packaged instructions that tell an agent how to perform a recurring task — have become a common way to extend agents such as coding assistants, and vendors including Anthropic have shipped skill systems that let teams add and share them. The study’s distinction is important for how those are built: if the benefit comes from imposing a reliable workflow rather than injecting new facts, then a skill’s value lies in its structure and scoping, not in how much information it carries. That reframes skill design as a retrieval and orchestration problem rather than a knowledge problem.
Technical Details
The researchers attribute most of the measured gains to the way skills constrain an agent’s steps into a repeatable procedure, rather than to new domain knowledge the skill supplies. The identified bottleneck is selection: past a certain library size, choosing the correct skill degrades performance, because the agent must first find the right tool among many before it can benefit from it — a variant of the context-management problems that appear as an agent’s available tools multiply. That points to concrete mitigations: tighter skill descriptions, hierarchical grouping, and retrieval over the library, rather than simply adding more skills.
Who’s Affected
Developers building on agent frameworks that support skills, including coding agents, are the most direct audience. Platform vendors that encourage large marketplaces of shareable skills face the study’s caution that scale can hurt as well as help, because a bigger catalog raises the selection burden. Teams running agents in production gain a specific variable — library size and selection quality — to monitor.
What’s Next
The practical follow-up is engineering better skill retrieval and grouping so that a large library remains usable, and measuring selection accuracy as libraries grow. The study frames skill selection, not skill quantity, as the next problem to solve for agent builders.
Related Reading
- Menlo’s Matt Murphy Says Anthropic Hit a $47B Run Rate by May
- Anthropic Blocks Chinese Firms From Claude Code as Alibaba Bans It Internally
- Researcher Used Claude to Uncover a Ticketing Flaw at Most US Festivals
- Anthropic Launches Claude Science, a Flagship AI Product for Research
- Claude Sonnet 4.8 Was Just Leaked in Anthropic’s Own Code
