Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees
This work develops a mathematical framework, Best Prefix Selection (BPS), to optimally choose a subset of skills for an LLM agent's context window. It treats skill selection as maximizing capability benefit while minimizing token cost under a hard budget. The method provides a provable bicriteria (1 - 1/e, 1) guarantee, ensuring the selected skills offer near-optimal performance relative to their size.