A study from researchers at Princeton and UC San Diego, covered by The Decoder, looked at why giving AI agents packaged "skills," reusable instructions for a specific task, improves their performance. The answer is not the one most teams assume when they start building an agent stack.
The finding that cuts against intuition
The improvement from adding skills comes mainly from structure, a defined sequence of steps for the agent to follow, rather than from the extra knowledge packed into the skill itself. An agent given a well-organized procedure for a task tends to outperform one given the same underlying facts as a loose block of text, even when the raw information is identical. Structure is doing more of the work than content.
That is useful on its own. Teams building agent workflows have been treating skills mainly as a knowledge problem, writing longer and more detailed instructions in the hope that more context produces better results. The study points at a different lever: how that context is organized matters more than how much of it there is.
Where it breaks down
The same research found a second effect that matters just as much for anyone scaling past a handful of skills. As the skill library grows, agents get measurably worse at finding and selecting the right one for a given task. More available skills does not mean better coverage. Past a certain library size, it means more time and more errors spent on the retrieval step before the agent even starts the task it was supposed to do.
This is the part that does not show up in a demo with three or four skills loaded and looks completely different once a team has built forty of them for every edge case they have ever hit.
What this means for building an agent stack
Two decisions follow directly from this, and both run against the instinct to add more.
Write skills as structured procedures, not reference documents. A skill that reads like a checklist with explicit steps will outperform one that reads like a paragraph of context, even when both describe the same task, because the structure is what the agent is actually using.
Curate the skill library on purpose instead of letting it grow by accretion. Every new skill added to a shared library is a small tax on every future retrieval, not a free addition. Splitting a large general-purpose agent into a few narrower ones, each with a small, well-scoped set of skills, avoids the exact failure mode this study documents, and it is a cheaper fix than trying to build a better retrieval layer on top of one bloated library.
The rule worth keeping
More skills is not the same as a more capable agent, and past some threshold it is a measurably less capable one. The teams getting real value out of agent workflows right now are treating the skill library the way a good engineer treats a codebase, worth pruning, not just worth adding to. A structured, narrow skill set an agent can reliably navigate beats a comprehensive one it gets lost in.
Building out an agent workflow and not sure which parts of your current AI stack are pulling weight versus adding overhead? AI Stack Optimizer benchmarks your actual tasks against the models and tools you are already paying for.