Anyone integrating AI agents into development or business processes will sooner or later work with so-called skills. This refers to stored instructions that tell an agent how to complete a specific task – from creating a ticket to the deployment check. A study by researchers at Princeton University and the University of California, San Diego now provides an important clarification: the benefit of such skills arises primarily from the structure they contain, not from additional specialist knowledge.

The difference between knowledge and procedure

The distinction sounds academic, but it has immediate consequences for practice. Many teams treat skills like a knowledge base: they write into them what the model should know about a system, an interface or a product. The study suggests that this is precisely not the decisive lever. Foundation models already bring a great deal of general knowledge with them. What they lack is the reliable sequence – which step comes first, which check precedes the next step, when a task counts as complete.

A skill therefore acts less like a reference work and more like a work instruction. It reduces the number of decisions an agent has to make itself, and thus lowers the probability that it chooses a path that sounds plausible but leads to the wrong result. For the architecture of agent setups, this means: the effort belongs in the description of the procedure, not in the most complete possible collection of background information.

The problem begins with the growth of the library

The second finding of the study is the more uncomfortable one. Skills only unfold their benefit if the agent also finds the right instruction at the right moment. As the skill library grows, this becomes increasingly difficult for it. The effect is known from other contexts: the more similar options are available to choose from, the less reliable the selection itself becomes.

As a consequence, the benefit tips over. A setup that works very well with a handful of clearly delineated skills can perform worse after expansion to several dozen instructions than before – not because the individual skills have become worse, but because the agent picks the wrong one or none at all. Anyone running agents in customer projects should plan for this point before the library has grown organically.

What this means for building agent setups

Some practical guidelines can be derived from the two findings:

  • Procedure before content: A skill should primarily describe steps, sequences and abort conditions. Background knowledge belongs in it only if it is really necessary for executing the step.
  • Clear delineation: Two skills that overlap in content increase the error rate in selection. A single skill with a branch is better than two almost identical variants.
  • Library size as a design parameter: The number of skills is not a mere matter of diligence, but an architectural decision. It is worth deliberately setting an upper limit and regularly pruning.
  • Relieve the selection: Instead of offering all skills at once, you can narrow the context – for example by domain, project phase or role. The agent then chooses from a few relevant options instead of the entire inventory.
  • Measure instead of assume: Whether a new skill improves operations can only be assessed with repeatable test cases. Without this safeguard, every extension remains a risk.

Parallels to classic process documentation

Anyone who has ever maintained a quality management manual or a collection of internal work instructions will recognise the pattern. There too, the rule applies: a lean, well-maintained documentation is used, an sprawling one is not. There too, application rarely fails because of missing knowledge, but because of the question of which document applies to the case at hand. Skills for AI agents apparently follow the same logic – with the difference that the selection is now automated and its errors land directly in the result.

This relieves teams in one place and burdens them in another. The relief is that skills do not have to be encyclopaedic; short, precise procedure descriptions are sufficient. Effort, by contrast, arises in maintenance: anyone creating skills needs a process that also merges, revises or deletes them again.

Classification for web and software projects

For projects in which agents take over recurring tasks – content migrations, routine reviews, evaluations, approval steps – the focus thus shifts. Success depends less on how much you teach the agent than on how clearly the procedures are cut and how disciplined the skill library is managed. It is advisable to plan agent setups from the outset with a manageable number of sharply delineated skills, to restrict access depending on context, and to review extensions just as you would code changes. Anyone who ignores this builds a system that convinces in pilot operation and quietly becomes unreliable in regular operation.

Sources