SkillsInjector: Dynamic Skill Context Construction for LLM Agents

SkillsInjector introduces a dynamic method for optimizing how large language model agents access and utilize skill libraries. Rather than treating skill selection as static, the approach adaptively determines which skills to include, how many to present, and how to describe them based on task requirements, achieving measurable performance improvements across multiple benchmarks.

SkillsInjector addresses a fundamental inefficiency in current LLM agent architectures. As skill libraries expand, developers face a counterintuitive problem: adding more capabilities can reduce performance rather than enhance it. This degradation stems from context pollution—irrelevant skills create noise that distracts the model from optimal task execution. The research validates what practitioners have observed empirically: static skill injection is suboptimal.

The two-stage approach reflects deeper architectural thinking about LLM behavior. The context planner learns execution-grounded preferences, meaning it understands which skills actually matter for specific task contexts rather than applying generic selection criteria. The adaptive budgeting component acknowledges that different tasks require different skill set sizes, eliminating the inefficiency of fixed-size skill presentations. The set-aware renderer represents the most sophisticated contribution—it contextualizes skill descriptions based on neighboring skills, recognizing that information value depends on what else is available.

These improvements (3.9–7.3 percentage points across benchmarks) matter because skill-augmented agents increasingly power production systems handling complex workflows. In enterprise environments, higher completion rates directly translate to reduced error handling, faster task resolution, and lower operational costs. The modular gains from ablation studies confirm each component contributes meaningfully.

The broader implication shapes how AI systems will scale skill libraries. Rather than linear addition of new capabilities, frameworks will adopt context-aware injection strategies. This pattern extends beyond skills to any injected context—retrieval-augmented generation, tool selection, memory management. Organizations building AI agent platforms will increasingly adopt dynamic context optimization as a competitive requirement, not an optional enhancement.

• →SkillsInjector dynamically selects, budgets, and presents skills to LLM agents rather than using static injection methods.
• →Performance improves 3.9–7.3 percentage points across tau2-bench, SkillsBench, and ALFWorld benchmarks compared to strongest baselines.
• →Adaptive skill budgeting recognizes that different tasks require different numbers of accessible skills for optimal performance.
• →Set-aware rendering tailors skill descriptions based on co-injected neighbors, treating context optimization as a relational problem.
• →This work shifts skill augmentation from binary selection to continuous optimization of injected context itself.