SZ SkillZip Pro
01 / 17
Research paper, explained

Compress the skill.
Keep the behaviour.

SkillZip Pro is an evaluation-free way to shrink evolving agent skill bundles while preserving how they load, what they can reach, and which entries can run on their own.

Execution-aware dynamic compression for progressively loaded skills
Bai et al., Alibaba Group and Zhejiang University · arXiv:2608.30785v1 · 31 August 2026
SKILL.md
route + contract
references
subskills
schemas
scripts + assets
01 · The short version

A production skill is a folder, not a prompt.

As an agent evolves, useful instructions get copied across roots, references and branches. SkillZip Pro removes the repetition, but keeps each branch reachable and each public entry usable.

What it does

Builds a typed map of the bundle, finds safe reuse, and emits an ordinary directory the existing harness can read.

Why it helps

Less shipped content, less repeatedly loaded context, and fewer tokens spent re-reading the same rules.

Why it is safer

Every removal needs evidence. Routing, contracts, locked files and the materialised output are audited before publication.

Useful for

Self-evolving research agents, tool-use skills, content moderation, spreadsheet agents, coding assistants, and any agent whose knowledge is split across files and branches.

02 · The beginner mental model

Agents do not load the whole bundle at once.

They progressively disclose it: discover the skill, activate the root, then open only the resources needed by the current task.

That means a token saved from an on-demand reference is not equivalent to a token saved from the root. The root is paid for repeatedly.

Catalog
Can the agent find it?Name, description and entry metadata before activation.
Activation
What is always read?The root instructions when the skill is selected.
Path
What does this task need?The transitive files loaded for one execution branch.
Deployment
What is shipped?Every byte in the bundle, including rarely used resources.
03 · The trap

A big ratio can hide a broken skill.

The paper tests three tempting shortcuts. Each optimises one view of the bundle while damaging another.

Root-only compression

Short root, dead branches

Rewriting only SKILL.md can remove the routing table. Files remain on disk, but the agent can no longer discover them.

Flat concatenation

Huge ratio, lost knowledge

Concatenate all files, compress the text, then map it back. Progressive-loading boundaries disappear, and branch-specific content can be deleted.

Global deduplication

Less storage, more runtime

Move shared text into the root and every task pays for it. A rare branch can become part of every activation.

SkillZip Pro

Share at the right scope

Factor repetition into the lowest safe activation scope, move long guarded sections into capsules, and keep route contracts intact.

04 · The method

Two pillars make compression useful.

1

Activation-aware cross-file compression

Find savings across the resource graph without promoting rare knowledge into the root.

  • Remove text already guaranteed by a signed host contract.
  • Share exact rules inside the scope that actually needs them.
  • Move long, explicitly guarded sections into on-demand capsules.
2

Cross-file routing preservation

Treat navigation as a contract, not ordinary prose.

  • Lock routing instructions and reference-bearing lines.
  • Keep relative paths and public entry points discoverable.
  • Re-read the emitted directory and reject unsafe candidates.
typed resource graphcandidate transformationscost-aware selectiondisk auditpublish or fallback
05 · The accounting

One compression ratio is not enough.

SkillZip Pro reports catalog, activation, path and deployment cost separately, then optimises a weighted objective.

Key intuition

Moving 500 tokens out of an on-demand file can be good. Moving those same 500 tokens into the always-loaded root can make every task worse.

Catalog
find
Activation
repeat
One path
task
Deployment
ship

The paper uses a default storage weight λ = 0.05, but reports all layers even when optimising the combined objective.

06 · Deployment choices

Choose when to compress, then choose where the result lives.

The axes are independent. Click a mode to see the paper's intended use.

When
Where
Persistent
Transient
One-Shot
One-Shot + Persistent

Smallest shipped bundle and lower steady-state cost.

One-Shot + Transient

Canonical bundle stays byte-identical.

Continual
Continual + Persistent

Zip-on-Write with periodic global repack.

Continual + Transient

Cached task-specific views for moving targets.

Initial migration or release packaging. Rewrite the canonical bundle once, run a multi-entry audit, then ship the smaller bundle for repeated use.
07 · The safety boundary

Evidence, not deletion count, sets the safe ceiling.

Each removal carries one witness. Stronger witnesses allow more savings, but the paper keeps interface contracts atomic and rejects anything it cannot justify.

W1

Literal containment. The removed text survives byte-for-byte at another reachable location.

W2

Coverage witness. A merged or shared representation preserves every protected literal and polarity.

W3

Entailment witness. A frozen checker confirms the decision-relevant content is already expressed elsewhere.

graph closure
scope + guards
public entries
locked bytes
contiguous contracts

No environment contract, no host-entailment deletion. If an audit fails, the method falls back to a verbatim bundle or raw execution closure.

08 · The compiler loop

It is a compiler with a hard stop.

The optimiser can propose aggressive changes. The disk-level audit decides whether anything is allowed to ship.

1. Resolve safe graph2. Extract typed contracts3. Estimate path weights4. Generate candidates5. Select under cost objective6. Materialise directory7. Audit from disk8. Atomic publish
If it wins and passes

Publish the ordinary relative-file directory and save the compressor state for the next evolution patch.

If it inflates or fails

Keep the patched source or raw closure. Compression is allowed to lose savings, not to lose a valid update.

09 · What the experiments found

SkillZip Pro was the only compressor that cut all four costs and kept quality.

+19.7%lower shipped tokens
+7.3%lower always-loaded tokens
+14.3%lower mean path tokens
+10.8%lower worst-path tokens
0.480 pooled
keeps ✓
reference
baseline
0.441
fails
0.382
lossy
Main cost figures: Table III. Pooled quality result: Table IV. The “keeps” criterion is the paper’s pre-set 5-point non-inferiority margin across 102 held-out tasks.
10 · The compounding effect

Self-evolution manufactures the redundancy it can later remove.

Across the paper’s evolved libraries, about 55% of each new round repeated text already written. A few universal rules absorbed most of that reuse.

29% → 53%

Share of shipped tokens removed from round 2 to round 15.

47%

Reuse captured by the top two of seven shared modules.

Compression savings compound with ageShare of shipped tokens removed rises from 29 percent at round 2 to 53 percent at round 15. 29%53%round 2round 15shipped tokens removed
11 · Production evidence
38.1%
less deployed content on a production content-moderation skill

The paper reports 10.4% lower pooled, task-paired runtime tokens over 200 audits, with 89% accuracy in the compressed runs.

ConfigurationDeploymentRuntimeAccuracy
Uncompressedreferencereference88-92%
Pro + W1/W213.8% lower7.1% lower89%
Pro + W332.7% lower6.8% lower88%
Pro + W3 inside root38.1% lower10.4% lower89%

The unprotected 71.4-75.8% configurations lost 18-26 accuracy points, primarily through additional false positives.

12 · Trade-offs

The right lifecycle depends on how the bundle changes.

Stable bundle, many calls

Persistent compression pays its build cost once and stays cheaper across repeated direct calls.

Frequent edits, changing views

Transient compression can rebuild only the affected execution view and preserve the canonical bundle.

Frequent edits, re-shipped library

Continual persistent compression uses local repair plus triggered global repacks to control drift.

Structural, not universal

The guarantee protects explicit contracts and routes. It does not prove every possible model interpretation remains equivalent.

Routing gets more important

Shared modules reduce duplication, but the executor must correctly follow the extra links. Cross-model results were mixed.

Evidence is a constraint

Without a stable, typed environment contract, host-entailment pruning is disabled. Unclear deletions are refused.

13 · How to implement it

Start with the bundle you already have.

Use the stepper as a practical implementation plan. The exact paper describes the compiler; this sequence turns it into an engineering rollout.

14 · Before you ship

Make the audit the product boundary.

If the candidate is smaller but a route, contract, locked byte or patch disappears, it is not a valid output.

Recommended first milestone

Build a read-only analyser that emits the resource graph, entry contracts, candidate witnesses and four cost layers before it is allowed to rewrite anything.

Every local reference resolves inside the bundle root.
Every public entry remains discoverable and independently usable.
Every removal has exactly one typed witness.
Interface contracts remain contiguous and verbatim.
Code, data, schemas and assets keep byte identity.
The output is audited from disk, not trusted from memory.
A failed continual run publishes the valid patched source.
Runtime savings are measured in paired execution sessions.
15 · The takeaway

Good compression understands the shape of execution.

SkillZip Pro’s contribution is not simply shorter text. It treats an agent skill as a living, routed resource graph, and makes every saving answerable to where and how the agent will load it.

Compress across files.

Remove repeated knowledge once, at the lowest safe scope.

Preserve the route.

Reachability and entry independence are first-class constraints.

Measure the run.

Deployment savings are useful only when execution cost and quality hold.

Reference notes

What this deck preserves from the paper.

Scope

This is an explanatory visual summary, not a reproduction of the formal proofs, all tables, or the full appendix.

Evidence

Reported figures are taken from the supplied PDF, especially Tables II-III, IV-VI, IX-XIII and XXIII.

Interpretation

Use cases and implementation framing are explanatory extensions. They are not additional empirical claims from the paper.

Prepared from the supplied file: “SkillZip Pro-Execution-AwareDynamicCompression of ProgressivelyLoadedSkills for Self-EvolvingAgents.pdf”.