SkillBrew: Multi-Objective Curation of Skill Banks for LLM Agents

SkillBrew introduces a multi-objective curation framework for managing skill banks in LLM agents, addressing the problem of bloated repositories filled with redundant and outdated skills. The approach treats skill bank management as a constrained optimization problem balancing utility, diversity, and query coverage, evaluated successfully on public benchmarks.

SkillBrew addresses a fundamental inefficiency in retrieval-augmented LLM systems: the tendency to accumulate skills without pruning. As LLM agents become more sophisticated and deployed in production environments, the quality of their decision-making depends heavily on the curated knowledge they access. This research reframes skill bank management from a passive append-only process into an active curation challenge, treating it as a Pareto optimization problem where multiple objectives must be balanced simultaneously.

The framework emerges from growing recognition that larger skill banks don't necessarily improve agent performance—redundancy and noise actually degrade decision quality and increase computational overhead. By formalizing curation as multi-objective optimization with a bi-level propose-then-verify architecture, SkillBrew provides a principled alternative to ad-hoc skill management. This mirrors broader trends in machine learning toward quality-over-quantity approaches and automated data curation.

For developers building LLM agents, this research suggests that maintaining skill banks requires active maintenance similar to database administration. Organizations deploying retrieval-augmented agents could improve performance by implementing systematic curation rather than continuously expanding repositories. The work has implications for agent efficiency, cost reduction, and reliability in production systems.

Looking forward, the automation of skill bank curation could become a standard component of LLM agent infrastructure. Future work might explore how curation frameworks adapt to distribution shifts, incorporate user feedback, or scale to enterprise-level skill repositories. The research positions intelligent knowledge management as critical to next-generation autonomous agent systems.

• →Skill bank curation framed as Pareto-aware multi-objective optimization rather than append-only accumulation
• →Redundant and outdated skills in repositories degrade agent performance and efficiency
• →SkillBrew's bi-level propose-then-verify loop balances utility, diversity, and query coverage
• →Active curation of knowledge bases becomes essential infrastructure for production LLM agents
• →Framework demonstrates that quality-controlled skill banks outperform ever-growing unmanaged repositories