thaki-AI/daily-paper-2026-07-17-skill-graveyard-safe-deprecation

Skill Graveyard: Safe Autonomous Deprecation in Growing Agent Skill Ecosystems

TL;DR — A telemetry-plus-redundancy autonomous deprecation policy with an actuation-time irreplaceability guard removes 7.39% of a production skill corpus (2,164 units) without measurable routing regression, while formalizing retrieval degradation under corpus growth as skill half-life.

ThakiCloud AI Research · 2026-07-17 · 📝 Tech blog (KO)

Problem

Production agent skill corpora degrade in routing precision as they grow (the SRA Noise Problem), and naive autonomous deprecation risks removing rare-but-irreplaceable skills (false deprecations).

Approach

We formalize skill half-life as the corpus size at which a retrieval metric falls to half its small-corpus value, then deploy a two-signal deprecation policy (usage telemetry + semantic redundancy clustering) with an actuation-time irreplaceability guard that re-simulates retrieval on a labeled suite before removing any skill.

Key contributions

• Skill half-life formalization: Top-1 accuracy collapses 42.2pp (0.778 to 0.356) as corpus grows 21.6x (100 to 2,164 units), with the half-life crossing within the current production range, while gated Recall@5 remains robust.
• Guarded deprecation policy: On a 63-case hand-labeled suite, a 7.39% corpus reduction achieves zero measured regression on every metric (Recall@5, gated Recall@5, Top-1, hallucination rate, negative-avoidance rate) while eliminating all 4 false deprecations an unguarded policy would commit.
• Generalizable signals: Usage telemetry and semantic redundancy do not depend on repository-specific features, making the policy applicable to any growing skill, plugin, or tool ecosystem including MCP server marketplaces.

Figures

Top-1 accuracy collapses 42.2pp as corpus grows 21.6x, while gated Recall@5 stays robust. (Measured on ThakiCloud production Claude Code harness (63-case suite). Stochastic subsampling at each corpus size.)
_{Measured on ThakiCloud production Claude Code harness (63-case suite). Stochastic subsampling at each corpus size.}

Guarded policy eliminates all 4 false deprecations at identical 7.39% corpus reduction. (Without safety guard: 4 false deprecations. With safety guard: 0. Corpus reduction identical at 7.39% (160/2164 skills) in both conditions.)
_{Without safety guard: 4 false deprecations. With safety guard: 0. Corpus reduction identical at 7.39% (160/2164 skills) in both conditions.}

Results (as argued)

Guarded removal of 160/2,164 skills (7.39%) produces identical routing metrics (Recall@5=0.822, gated Recall@5=0.667, Top-1=0.378, hallucination=0.000, negative-avoidance=0.375) as the full corpus on the 63-case suite. Without the guard, 4 false deprecations occur. The approach is regression-free-by-construction on the guarded suite but regression-oracle-relative, not fully independent held-out evidence.

Limitations

Single-repository case study; zero-regression claim is bounded by the 63-case labeled suite (continuous-monitoring evaluation is the essential next step). No random-removal or usage-only ablation. Guard mechanism verified only in aggregate counts, not per-cluster logs. Negative-avoidance rate (0.375) is a pre-existing harness weakness unaffected by pruning.

Abstract

Production agent harnesses increasingly route user queries across large, ever-growing corpora of reusable "skills" (prompts, subagents, and tool definitions). As the corpus grows, routing precision degrades -- the "SRA Noise Problem." Our two prior technical reports diagnosed this degradation and partially repaired it by improving the retriever and query decomposition, but left the retriever ceiling well below saturation. This paper pivots to the opposite lever: instead of improving the retriever, we ask whether shrinking the corpus itself, through safe autonomous deprecation of redundant and dead skills, can remove low-value skills without regressing routing precision -- and what safeguard prevents the corpus from losing rare-but-irreplaceable skills. We formalize retrieval degradation under corpus growth as "skill half-life," measure it on ThakiCloud's own production Claude Code harness (2,164 registered units, 1,930 skills), and show that Top-1 accuracy collapses 42.2 percentage points as the corpus grows 21.6x, while an abstain gate is comparatively robust. We then present a two-signal (usage telemetry plus semantic redundancy) deprecation policy with an actuation-time irreplaceability guard. On a 63-case hand-labeled suite, a guarded 7.39% corpus reduction produces zero measured regression on every metric while eliminating all four false deprecations an unguarded policy would commit. We emphasize that this run measures no precision gain; reversing the growth-curve degradation is future work. Because the same 63-case suite drives the safety guard, the zero-regression re

Files

• LaTeX source
• References (BibTeX)

Citation

@techreport{thaki_skill_graveyard_safe_deprecation_2026,
  title  = {Skill Graveyard: Safe Autonomous Deprecation in Growing Agent Skill Ecosystems},
  author = {ThakiCloud AI Research (Hyojung Han)},
  year   = {2026},
  institution = {ThakiCloud}, note = {thaki-AI/daily-paper-2026-07-17-skill-graveyard-safe-deprecation}
}

---

Generated by ThakiCloud nightly research pipeline. License: CC BY 4.0.