Researchers at Princeton University and UC San Diego have identified a critical paradox in how AI agents use pre-built "skills" to complete tasks. Their findings reveal that skills help agents primarily by organizing workflows into structured sequences, not by providing additional knowledge. Yet this benefit collapses as skill libraries expand, creating a selection problem that degrades agent performance.
The study examined how large language model-based agents interact with skill libraries—collections of pre-programmed functions or instruction sequences designed to help agents solve complex problems. The conventional assumption held that skills work by encoding domain knowledge that agents could tap into. The research shows this assumption is wrong.
Instead, skills function as structured scaffolding. They force agents to think through problems in organized steps rather than attempting to reason through tasks in unstructured ways. This workflow organization makes agents more reliable and reduces hallucinations. When an agent has access to a well-designed skill like "write to file" or "search database," it doesn't gain hidden knowledge. It gains a reliable path to complete intermediate steps that would otherwise confuse or overwhelm the language model.
But the research exposes a hard limitation. As skill libraries grow from small collections to dozens or hundreds of options, agents face an expanding search problem. They must select which skills to use for each task from an increasingly large menu. Language models struggle with this selection at scale. They cannot reliably identify which five skills, from a library of fifty, matter for a given problem. This creates a bottleneck where more skills produce worse performance, not better.
The implications reshape how teams should design AI agent systems. Building massive skill libraries offers no advantage if agents cannot navigate them. Instead, the research suggests that curating smaller, task-specific skill sets produces better results. Organizations need to think about skill libraries as human-readable hierarchies rather than flat collections. Agents need structured ways to search through available tools.
This finding also explains why many deployed AI agents perform worse with time. As teams add new skills to improve coverage, they often degrade agent performance on existing tasks by creating selection noise. The agent spends computational effort evaluating irrelevant options.
The research carries implications for retrieval-augmented generation (RAG) systems and broader AI infrastructure. If agents struggle to select from large skill libraries, they likely struggle similarly when choosing from massive document collections or knowledge bases. The bottleneck is not memory or reasoning power. It is selection among abundant options.
For practitioners, the findings suggest several design principles. First, segment skill libraries by task type or domain rather than creating unified collections. Second, implement explicit skill filtering mechanisms that reduce the agent's search space before selection begins. Third, test agent performance with growing libraries and accept that curation beats accumulation.
The Princeton and UC San Diego teams have created a more precise picture of AI agent limitations. Skills work, but not because they store knowledge. They work because they structure thinking. And that structure breaks down when presented with too many options to choose from.