More Skills, Worse Agents? Skill Shadowing Degrades Performance When Expanding Skill Libraries

Expanding the libraries of skills available to large language model agents can degrade performance by up to 21%, according to a new study, which identifies the failure to select the correct skill—not the enlarged context—as the primary bottleneck [1][2]. The study, posted to arXiv on 21 May 2026, examines how skill libraries allow LLM agents to load task-specific instructions on demand, enabling non-expert users to solve domain-specific tasks through natural language without knowing which skills exist [1][2]. The researchers formulated the performance degradation as the pass rate drop between loading a library of known-helpful skills and a full 202-skill library [2]. To isolate the cause, the authors decomposed the pass rate drop into two effects: skill shadowing, where the agent selects wrong skills more often as the library expands, and context overhead, where the enlarged context degrades execution even when selection is correct [2]. Empirical estimates showed that the skill shadowing effect grows with library size and significantly contributes to the performance degradation, whereas the context overhead effect remains small and indistinguishable from zero [2]. LLM agents are a class of AI systems built on large language models—generative pre-trained transformers that generate text, speech, and images in response to user prompts [3]. OpenAI's ChatGPT, released in November 2022, accelerated the AI boom and reached 100 million monthly active users within two months [3]. The chatbot has been lauded for its potential to transform professional fields, but also criticized for generating plausible-sounding but incorrect answers, known as hallucinations, and for biases in its training data [3]. The study's finding that skill selection failure is the primary bottleneck when expanding skill libraries carries implications for the design of agent systems that rely on modular, on-demand instructions [2]. The authors derived upper bounds on both skill shadowing and context overhead to characterize their magnitudes of impact on the pass rate drop [2]. The observed asymmetry between the two effects suggests that efforts to improve agent performance with larger skill libraries should focus on refining the skill selection mechanism rather than on reducing context size [2].