thaki-AI/daily-paper-2026-07-31-skill-birth-synthesis-boundary

Skill Birth: Calibrating the Synthesis-versus-Reuse Decision Boundary in Self-Evolving Agent Skill Ecosystems

TL;DR — Skill provenance (individual authoring vs. bulk import) barely predicts whether a skill is ever reused (39.0% vs 35.1%), but strongly predicts how fast and how intensively it is reused: 24x faster time-to-first-reuse (1.2 vs 29.0 days median) and 2x higher post-birth modification count (1.11 vs 0.54). The synthesis-versus-reuse decision should be based on reuse tempo, not binary occurrence.

ThakiCloud AI Research · 2026-07-31 · 📝 Tech blog (KO)

Problem

When a skill router abstains, the harness must choose between native fallback, composing existing skills, or synthesizing a new skill — a decision that prior work has left implicit. Existing literature evaluates synthesis mechanisms at a point in time on benchmarks, but none measures what happens to synthesized skills over time in a real production corpus.

Approach

Audit of full git history of the skill directory of a production self-evolving agent skill ecosystem: 811 commits touching that directory, yielding 1,972 distinct skill birth events. Each birth is assigned to one of five provenance cohorts (bulk_import, pipeline_autonomous, sync_import, deliberate_feat, other_singleton) by birth batch size and commit-message regex. Four reuse survival measures are computed per cohort: ever-reused rate, never-reused rate, median days to first reuse (conditional), and mean post-birth modification count.

Key contributions

• First empirical characterization of the synthesis-versus-reuse decision boundary via birth-to-reuse survival analysis across five provenance cohorts in a real production ecosystem (811 skill-directory commits).
• Skill-Birth Decision Policy (SBP) over four axes — recurrence, coverage by composition, batch opportunity, and authorship channel — retrospectively grounded in measured survival data.
• Counter-intuitive finding: fully autonomous synthesis (pipeline_autonomous, n=30) shows the second-lowest abandonment rate and by-far highest reuse intensity (2.83 mean mods), suggesting under-deployment rather than the commonly assumed over-synthesis risk.

Figures

Individually authored cohorts consistently outperform bulk import on ever-reused rate, with deliberate_feat achieving the highest at 45.1%. (Measured in a CPU-only container)
_{Measured in a CPU-only container}

Conditionally on reuse occurring, individually authored skills reach first reuse 24x faster than bulk-imported skills (1.2 vs 29.0 days median). (Measured in a CPU-only container)
_{Measured in a CPU-only container}

pipeline_autonomous skills accumulate the highest reuse intensity (2.83 mean modifications), more than double any other cohort, suggesting sustained engagement despite being the smallest cohort (n=30). (Measured in a CPU-only container)
_{Measured in a CPU-only container}

Results (as argued)

Individually authored skills reach first reuse in median 1.2 days vs 29.0 days for bulk import (24x faster) and accumulate 1.11 vs 0.54 mean post-birth modifications (2x more intensive), yet the binary ever-reused rate gap is only 39.0% vs 35.1%. Autonomous synthesis (n=30) has 56.7% never-reused rate (second-best) and 2.83 mean modifications (best). The prevailing concern about autonomous synthesis causing skill-sprawl is not borne out in this corpus.

Limitations

Single-repository, external validity limited to one organization's adoption choices and working rhythm. Cohort classifier is a lexical proxy (batch size + commit-message regex), not verified ground-truth authorship. Right-censoring deflates apparent reuse rates for recent births. Modification conflates genuine reuse-driven iteration with incidental touch-ups. Survivorship: capability gaps where no skill was ever attempted are structurally invisible to git archaeology.

Abstract

When a skill router abstains because no existing skill clears its retrieval gate, the harness faces a decision that prior work has largely left implicit: fall back to a native, skill-free response; compose two or more existing skills; or pay the authoring cost of synthesizing a brand-new skill. We study this synthesis-versus-reuse boundary empirically by auditing the full git history of the skill directory of a real, actively used production self-evolving agent skill ecosystem: 811 commits touching that directory, yielding 1,972 distinct skill birth events, where a birth is the first commit that adds a skill manifest. Each birth is assigned to one of five provenance cohorts (bulk import, autonomous pipeline synthesis, cross-repository sync, deliberate feature authoring, and unlabeled singleton addition) using birth batch size and commit-message structure, and each is followed forward in time to measure whether, how quickly, and how intensively it was later modified. Every quantity we report is a descriptive statistic computed over the complete population of 1,972 births; because there is no sampling frame, we report no inferential tests, confidence intervals, or p-values. The headline result is a decomposition: skills that were individually authored rather than bulk-imported reach first reuse roughly 24x faster in wall-clock terms (median 1.2 versus 29.0 days) and accumulate roughly 2x the post-birth modification count (1.11 versus 0.54), yet their raw ever-reused rate is only modestly higher (39.0% versus 35.1%). Speed and intensity of reuse, not the binary fact of reuse,

Files

• 📄 Paper (PDF)
• LaTeX source
• References (BibTeX)

Citation

@techreport{thaki_skill_birth_synthesis_boundary_2026,
  title  = {Skill Birth: Calibrating the Synthesis-versus-Reuse Decision Boundary in Self-Evolving Agent Skill Ecosystems},
  author = {ThakiCloud AI Research (Hyojung Han)},
  year   = {2026},
  institution = {ThakiCloud}, note = {thaki-AI/daily-paper-2026-07-31-skill-birth-synthesis-boundary}
}

---

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