22 August 2026

Study reveals why AI agents improve with instruction templates, and their limits

• Researchers from Princeton, UC San Diego, and other universities tested AI agents on 8,135 tasks to understand why pre-written instruction templates called skills help them work better.
• Skills work primarily by giving agents a reliable process to follow, not by teaching them new facts. This procedural structure accounted for 65.7 percent of performance gains.
• Skills create new failure modes: agents sometimes apply instructions too rigidly or use irrelevant templates, causing errors in about 10 percent of cases tested.
• As skill libraries grow larger, agents struggle to find the right template. Retrieval accuracy dropped from 29.6 percent with 5 skills to 3.3 percent with 100 skills.
• The researchers argue that simply collecting more skills will not improve agents. Instead, building better systems for creating, retrieving, and applying skills matters more.

Reported by The Decoder