SkillZip Pro: Execution-Aware Dynamic Compression of Progressively Loaded Skills for Self-Evolving Agents

2608.30785 ↗ · 2026-08-31 · cs.AI

Xiaofan Bai, Chao Liu, Hongqiang Lin, Di Wu, Mingli Song, Xuan Jin, Xipeng Cao, Yuhong Li

The finding

Headline claim

An evaluation-free, harness-agnostic compressor for progressively loaded agent skill bundles removes 38% of bundle tokens and 10.4% of end-to-end per-run tokens on a production content-moderation skill with no quality loss, while preserving routing and public entry independence.

The test

A cross-file, contract-preserving compression compiler (SkillZip Pro) was applied to production skill bundles and evaluated in an industrial multi-round harness with a full run matrix: compressed vs. reference conditions and flat prompt-compression baselines, measuring four cost layers (catalog, activation, path, deployment), task accuracy, routing preservation, and savings accumulation over continual skill-library evolution.

Effect size & scope

On the production content-moderation skill: 38% deployed-token reduction and 10.4% per-run token reduction with no accuracy loss; an aggressive 71% unprotected configuration lost up to 26 accuracy points. Gains compound across repeated use and growing libraries. Results are shown on essentially one production skill plus a multi-entry bundle; scope beyond content-moderation-style skills is untested in the provided text.

Limitations

The authors concede the structural faithfulness guarantee does not assert behavioral equivalence for every language model, protects only explicit contracts not latent readings, and requires author-declared entry contracts and conservative reference resolution (unrecognized files get no inferred dependencies). Unstated: single-skill main evidence, in-house harness not independently reproduced, environment-contract entailment depends on the audited environment digest holding.

Counterpoint

Industrial PR dressed as science: the headline number is one Alibaba production skill on the authors' own harness, with no external replication, no public task distribution details verifiable from the text, and a modest 10.4% end-to-end saving that may vanish under different traffic mixes; compression baselines were weak (flat compressors unsuited to bundles by construction).

Economy implication

Little direct evidence yet: if it generalizes, evaluation-free skill compression lowers inference cost for self-evolving agents at the margin, but one production case does not establish economy-wide inference-cost trends.

Who this matters for

Technical Capability: Agent builders deploying large skill libraries get a concrete mechanism to cut context cost without touching the harness, addressing the real failure mode where compression breaks routing.

Market & Valuation: If per-run token costs on agentic workloads drop ~10% for free at deployment time, inference margins for agent platforms improve slightly, though the evidence base is one internal skill.

Abstract

Production agent skills are directory bundles, not isolated prompts. The root is loaded at activation; references, schemas, scripts, assets, and nested subskills are loaded only when an execution path needs them. Compressing only the root misses most deployment cost and may move branch-specific details into the always-loaded context. Flattening instead destroys progressive-loading boundaries. We introduce \method, an evaluation-free compressor for complete, progressively loaded skill bundles. It leaves the agent harness unchanged and emits an ordinary directory. The method combines two safeguards. First, it compresses \emph{across files}, removing content from a reference or subskill when the root or a declared environment contract already provides it. Second, it preserves routing, so every required file and directly callable entry remains reachable after rewriting. Users can configure \method along two independent axes. \emph{One-Shot} mode rebuilds the full bundle; \emph{Continual} mode reuses state and applies Zip-on-Write after each evolution patch. \emph{Persistent} compression rewrites the shipped bundle to reduce storage and runtime context. \emph{Transient} compression keeps that bundle byte-identical and builds a task-specific view, reducing only per-run context after build cost. Entry contracts mark private, public, and conditional resources; a multi-entry audit preserves standalone public subskills. On a production content-moderation skill evaluated by our industrial multi-round harness, \method removes \hl{38\%} of skill bundle tokens and \hl{10.4\%} of end-to-end per-run tokens with no quality loss, while an unprotected 71\% configuration loses up to 26 accuracy points to one-sided false positives. On a multi-entry bundle, \method effeciently reduces token cost while near-perfectly preserving every route and public entry.