New Framework Optimizes AI Agent Skills for Software Tasks While Cutting Costs

A research team has developed a systematic approach to designing agent skill bundles—the instruction sets that configure AI systems for coding and software engineering work. Rather than treating skills as fixed artifacts, the new framework treats them as optimization targets, balancing multiple competing goals simultaneously.

The Problem with Static Skills#

Current practices rely on manually crafted or single-objective evolved skills, typically optimized for pass rate alone. This overlooks a critical tension: a skill configuration might boost task completion rates while simultaneously consuming far more computational resources, or it might provide guidance that sounds helpful but actually misleads the agent. The result is that deployed skill bundles often leave better configurations on the table.

How SkillMOO Works#

The framework, called SkillMOO, uses a two-stage process to explore the skill design space. It leverages large language models to propose edits to existing skill bundles, then applies NSGA-II Pareto selection—a multi-objective optimization technique—to identify skill configurations that trade off pass rate and inference cost. This approach surfaces the frontier of non-dominated solutions, where improvements in one dimension require sacrifices in another.

Empirical Results#

Testing across all 16 tasks in SkillsBench, a software engineering benchmark, SkillMOO claimed the top pass rate ranking on 11 of the 12 tasks with non-zero baseline performance. At the same time, the optimized skill bundles achieved cost reductions of up to 31.7% relative to static baselines, with pass rate improvements reaching as much as 21 percentage points on some tasks.

Key Insights for Skill Design#

Analysis of 38 successful skill edits revealed that pruning (removing unnecessary elements) and substitution (swapping one skill for another) were the dominant mechanisms of improvement. This finding suggests concrete design principles for practitioners building skill bundles without expensive search overhead.

Why It Matters#

The work highlights that deploying skills without cost-aware validation represents a missed opportunity. By framing skill engineering as a search problem with multiple objectives, the research opens a new direction: cost-aware, search-based approaches to configuring AI agents for practical software tasks.