New system helps small AI agents pick the right skills within budget constraints

Researchers have introduced a framework designed to help smaller language model agents intelligently select and combine skills from growing libraries while staying within cost and performance constraints.

The Challenge#

As skill libraries become more central to how LLM agents operate, the methods for choosing which skills to use have remained fairly basic—typically just ranking skills like documents and returning a fixed set of top options. This approach doesn't account for the real-world pressures that agents face, such as budget limits, quality requirements, skill dependencies, and the risk of errors.

How SkillSelect-Serve Works#

The new system, called SkillSelect-Serve, rethinks skill selection as a service recommendation and composition problem. Rather than treating skills as simple text items, it represents them as structured Skill Services that carry rich metadata: functional descriptions, dependencies on other skills, resource costs, error risks, and quality-of-service attributes.

The framework operates in two key steps:

• Planning phase: A local planner translates a task described in natural language into a structured set of service requirements
• Selection phase: A shared discovery system retrieves candidate skills from a large registry, then applies a dual-layer optimization process that evaluates both individual skill contributions and how well they work together as a bundle

The optimization stage balances multiple trade-offs—coverage (does the bundle handle the task?), redundancy (are some skills overlapping?), costs (does it fit the budget?), and risk (how likely are failures?).

Results and Implications#

When tested on over 35,000 skills and nearly 600 real task queries, SkillSelect-Serve delivered better performance than traditional top-k selection methods when compared at the same resource budget. The system improved both the recall of relevant skills and the overall utility of the selected bundles, suggesting that smarter composition can do more with limited resources—a key advantage for deploying agents with tight operational constraints.