Preprint
Reinforcement Learning

SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries

Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, Wenjie Zhang
August 6, 2026
arXivPDF

0

Citations

0

Influential Citations

—

Venue

2026

Year

Abstract

Large Language Models (LLMs) increasingly act as agents whose procedural knowledge is stored in reusable skill packages and loaded at inference time. As skill libraries grow, a central challenge is to expose the smallest sufficient executable context under a limited context budget. Existing systems struggle to reuse routines below the whole-skill level, preserve procedural contracts during compression, keep compressed routines executable and expandable, and update the compressed library as skills evolve. These challenges reveal a unit mismatch: skills are retrieved as packages, compressed as text, and converted into execution graphs only after retrieval, whereas reliable reuse requires a contract-bearing procedural unit. We propose SkillZip, an execution-aware procedural abstraction framework that performs contract-preserving compression over section-level graphs. SkillZip rewrites recurring contract-valid motifs into reversible ported macros while preserving boundary signatures, dependency closure, verifier reachability, and source-level expansion. At inference time, it hydrates a compact, dependency-closed context and expands macros only when required. ReZip further integrates new skills and revises risky macros using execution evidence. Comprehensive experiments1 on technical and embodied agent benchmarks show SkillZip consistently outperforms the strongest baseline by up to 12.2 points, while achieving a 3.46x compression ratio with 99.2% dependency preservation and 98.7% verifier reachability. Scaling analyses further confirm robust retrieval across skill libraries ranging from 200 to 100K skills.

Analysis

Why This Paper Matters

As LLM agents become more prevalent, the management of procedural knowledge—stored as reusable skill packages—has become a bottleneck. Existing systems retrieve skills as whole packages, compress them as text, and only convert to execution graphs after retrieval, leading to a unit mismatch that hampers reliable reuse under limited context budgets. SkillZip directly tackles this by introducing a contract-bearing procedural unit at the section level, enabling fine-grained compression without losing executability.

The paper's focus on contract preservation is particularly significant. In agentic systems, skills often have preconditions and effects (contracts) that must be maintained for correct execution. By preserving boundary signatures, dependency closure, and verifier reachability, SkillZip ensures that compressed routines remain valid and expandable, addressing a gap in prior compression methods that often sacrifice procedural integrity for size reduction.

Technical Contributions

  • Section-level graph compression: SkillZip operates on section-level graphs rather than whole-skill packages, allowing reuse of sub-skill routines.
  • Reversible ported macros: Recurring contract-valid motifs are rewritten into macros that can be expanded back to source-level, ensuring reversibility and maintainability.
  • Contract preservation: The framework maintains boundary signatures, dependency closure, and verifier reachability, which are critical for reliable execution.
  • ReZip for dynamic updates: New skills can be integrated and risky macros revised using execution evidence, making the library adaptive.
  • Inference-time hydration: Only the necessary context is loaded, and macros are expanded on demand, optimizing context usage.

Results

The experiments on technical and embodied agent benchmarks show that SkillZip consistently outperforms the strongest baseline by up to 12.2 points, indicating significant improvements in task success or efficiency. The compression ratio of 3.46x demonstrates substantial context savings, while the high dependency preservation (99.2%) and verifier reachability (98.7%) confirm that the compressed routines remain valid and executable. The scaling analysis from 200 to 100K skills shows robust retrieval, suggesting that SkillZip can handle large-scale skill libraries without degradation.

Significance

SkillZip introduces a new paradigm for skill library management in LLM agents, shifting from text-based compression to execution-aware graph abstraction. This could enable agents to handle more complex tasks with limited context windows, improving both efficiency and reliability. The contract-preserving approach also sets a standard for future work on procedural knowledge compression, potentially influencing areas like robotics, automated planning, and software engineering where reusable routines are essential. The ability to update the library dynamically with ReZip further enhances its practical applicability in evolving environments.

reinforcement-learningllm-agentsskill-librariesgraph-compressionprocedural-abstractioncontext-efficiency
Large Language ModelsRetrieval Augmented GenerationScaling LawsBenchmarks

Related Papers