SkillsInjector: Dynamic Skill Context Construction for LLM Agents

Abstract: LLM agents now draw on growing skill libraries to handle complex tasks. However, injecting more skills does not always improve task completion and can even degrade it. Existing methods still treat skill injection as a static step, selecting skills with fixed criteria, fixing the budget in advance, and leaving descriptions unchanged. We argue that this static treatment can undermine the utility of skills, because which skills are exposed, how many are included, and how they are presented all affect downstream performance. We propose SkillsInjector, a two-stage adaptive method that jointly addresses these decisions. First, a context planner learns execution-grounded skill preferences and admits an adaptive number of skills for each task. A set-aware renderer then tailors how selected descriptions are presented relative to their co-injected neighbors. Across tau2-bench, SkillsBench, and ALFWorld, SkillsInjector achieves the highest score, improving over the strongest baseline by 3.9, 6.1, and 7.3 percentage points, respectively. Ablation studies show that skill selection, adaptive budgeting, and set-aware rendering each contribute to the gain. These results show that skill-augmented agents benefit from optimizing the injected context itself. Code will be released upon publication

• The paper shows that static skill injection leads to performance collapse as candidate pool size grows, while dynamic injection with adaptive budgeting improves task success rates.
• It introduces a dual-stage pipeline (context planner and set-aware renderer) that optimizes skill selection and description rendering through execution-grounded utility predictions.
• Empirical evaluations demonstrate a 3.9–7.3 percentage point gain over retrieval-based methods, highlighting the effectiveness of adaptive context construction.

Dynamic Skill Injection for LLM Agents: SkillsInjector

Motivation and Challenge: Static Injection Bottlenecks

LLMs augmented with skill libraries exhibit scaling bottlenecks: increasing the injected skill pool often impairs rather than improves downstream task performance. Empirical analyses demonstrate a collapse in success rates as candidate pool size increases when skills are all injected statically, resulting from attention dispersion and skill competition. Surface-level selection methods (semantic similarity, fixed budgets, static descriptions) fail to address this, leaving agents overloaded or misaligned (Figure 1).

Figure 1: Scaling the candidate skill pool on tau2-bench. Static all-injection (red) collapses as candidate pool grows, whereas SkillsInjector (blue) remains stable.

These observations motivate a paradigm shift to task-adaptive skill context construction, requiring a dynamic injection mechanism that judiciously selects, budgets, and contextualizes skills for each task.

SkillsInjector: Adaptive Skill Context Construction

The paper introduces SkillsInjector: a two-stage pipeline resolving skill selection, set budgeting, and presentation jointly. The method comprises:

• Context Planner: Learns execution-grounded skill utility. Given a task and fixed skill library, it predicts the actual incremental task completion benefit ($\Delta(t,s)$) for each skill, using a combined loss (distribution alignment + pairwise preference) to separate fine-grained functional distinctions. Adaptive budgeting is performed per task by thresholding normalized predicted utilities.
• Set-aware Renderer: Further adapts selected skill descriptions by conditioning on task and injected skill set, clarifying boundaries and reducing harmful overlap. Renderer training leverages curriculum distillation from a strong teacher model, mixing trace-rich and trace-free supervision.

The pipeline yields an optimized skill context, tailored on both selection and description for maximal downstream utility, injected at inference into a fixed agent policy (Figure 2).

Figure 2: SkillsInjector pipeline integrating learned planner and set-aware renderer to inject adaptive skill sets and rendered contexts.

Core Contributions and Empirical Results

SkillsInjector achieves robust improvements across tau2-bench, SkillsBench, and ALFWorld benchmarks, outperforming classical retrieval, skill routing, and generative baselines by 3.9--7.3 percentage points in pass rate. Ablations confirm that skill selection, adaptive budgeting, and set-aware rendering each contribute independently to gains. The method remains stable as skill pool scales, mitigating collapse phenomena observed in static methods.

Key empirical findings include:

• Adaptive Budgeting: Per-task injected skill counts ($B_t$) show significant task-dependent variation, avoiding fixed-size saturation and capturing heterogeneous skill demand distributions (Figure 3).

  Figure 3: Histogram of per-task selected budget $B_t$, showing task-dependent distributions and occasional cap-bound spikes.

• Admission Threshold Robustness: Performance plateaus across a wide range of threshold values; both accept-all and accept-none extremes revert to degraded baselines, demonstrating the importance of intermediate, task-adaptive budgeting (Figure 4).

  Figure 4: Admission-threshold sweep across task domains, highlighting the operational plateau for task completion.

• Per-Skill and Per-Task Analysis: Skill utility is highly heterogeneous; many skills increase interaction cost, and a long tail actively reduces pass rate. Aggregating skills at domain level is suboptimal (Figure 5). Further, task responses to larger skill sets diverge strongly, with some tasks improving and others degrading with larger $N$ (Figure 6).

  Figure 5: Per-skill effects on pass rate and agent message cost, highlighting necessity of utility-based selection.

  

  Figure 6: Per-task responses to skill set scaling, motivating per-task adaptive budgeting.

Technical Design: Planner/Renderer Optimizations

The planner uses a dual-loss to ensure both global benefit alignment and local preference separation, relying on execution-derived labels—not just semantic signals—and hard negatives to discriminate functionally similar but unhelpful skills. Empirical evidence shows both losses are necessary; model capacity scaling yields negligible improvement compared to richer supervision. Renderer curriculum learning ensures set-aware contextualization, matching teacher performance at significantly reduced inference cost, and further reduces agent message overhead.

Implications and Future Directions

SkillsInjector reframes skill injection as a context optimization problem, emphasizing the conditional value of skills on task-specific and set-specific interplay. The approach highlights the limits of retrieval-only paradigms and demonstrates the importance of adaptive context construction for LLM agent systems. This architecture enables practical deployments where agent policies remain frozen and skill libraries fixed, achieving significant real-world gains in tool-centric LLM settings.

Potential theoretical extensions include co-training the skill library with the injection layer under shared utility objectives, enabling joint optimization of skill definitions and injection mechanisms. Practical next steps include evaluating in high-horizon environments (GUI automation, open web agents, etc.), diverse skill source pools, and proprietary model backbones.

Conclusion

SkillsInjector establishes dynamic skill context construction as a primary lever for agentic LLM performance, moving beyond static retrieval to task-adaptive skill selection, budgeting, and set-aware description rendering. Empirical results support substantial performance gains and resilience to scaling bottlenecks. The paradigm invites further research into joint agent-skill optimization and more complex interactive environments.