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

The Core Problem: Skills Don't Compress Well

When large language models act as autonomous agents, they need access to reusable skills—think of them as executable routines for tasks like navigation, object manipulation, or API calls. As skill libraries grow to thousands of skills, there's a hard constraint: the LLM's context window is finite. Existing systems compress skills by treating them as plain text, which destroys the executable structure that makes skills actually work. You lose the ability to reuse sub-routines, break contracts between skill components, or expand compressed skills later when needed.

The authors' insight is fundamental: skills are retrieved as packages but compressed as text, yet they need to work as executable graphs. This mismatch means compression breaks the procedural dependencies that keep skills valid and executable.

How SkillZip Works

SkillZip operates at the execution graph level—it models skills as directed graphs of operations and dependencies before compression happens. The key idea is to identify recurring motifs (patterns of operations that appear across multiple skills) and replace them with compact, reversible macros.

Critically, SkillZip preserves four contracts while doing this:

1. Boundary signatures: input/output types remain valid
2. Dependency closure: all operations a macro depends on stay reachable
3. Verifier reachability: constraints that validate skill correctness remain checkable
4. Source-level expansion: compressed macros can be expanded back to readable code if the LLM needs to debug or refine them

At inference time, the system loads a minimal, self-contained context and only expands macros on demand—you don't pay the full cost of uncompressed skills.

The authors also introduce ReZip, an update mechanism that integrates new skills into an existing compressed library and identifies risky macros using execution traces. When a macro fails, it gets flagged for revision.

Why It Matters: The Numbers

Experiments on technical and embodied agent benchmarks (the abstract doesn't name them, but "technical" suggests code/API tasks and "embodied" likely means robotics or simulation) show SkillZip achieves:

• 12.2 point improvement over the strongest baseline
• 3.46x compression ratio: skills shrink by nearly a factor of four
• 99.2% dependency preservation: macros maintain access to the operations they need
• 98.7% verifier reachability: validation still works after compression

The system scales to libraries with up to 100,000 skills, suggesting it avoids the combinatorial blow-up that naive compression would hit.

The practical payoff is clear: agents get much larger, richer skill libraries within the same context budget. The agent can call more specific, fine-grained routines instead of working with coarse-grained skills. And because macros are reversible and verifiable, skills remain debuggable and trustworthy—essential when agents operate autonomously.