"How your agents can write and optimize their own skills"

Why this is in the vault

This issue is a direct dispatch from the frontier of harness-engineering: automated text-space optimizers (SkillOpt, GEPA, EvoSkill) that treat .md skill files as trainable parameters — precisely the surface RDCO manages manually today.

⚠️ Sponsorship

Sponsored by TimescaleDB (Tiger Data). One item at the top of the issue ("From Tiger Data") promotes a CERN/TimescaleDB webinar on time-series analytics at scale, and a dedicated sponsor panel appears mid-issue. Editorial content (the Sunday Deep Dive) is independent of the sponsor.

Issue contents

Sunday Deep Dive — authored by Ben Dickson (TechCrunch / VentureBeat) Third-party content; Dickson is a contributing author, not AlphaSignal staff.

The piece frames the current state of skill-file optimization and surveys three emerging frameworks:

• SkillOpt (Microsoft Research) — treats skill .md files like neural network weights. Pipeline: rollout → evaluate → reflect → bounded edits. A "textual learning-rate budget" caps edit scope per iteration to prevent regressions. Validated on a held-out set; failed edits go to a rejected-edit buffer. Reported +23.5 / +24.8 point average improvement on GPT-5.5 across 52 harness/model/benchmark combos. Optimized files median ~920 tokens and shown to generalize across models.

• GEPA (Genetic-Pareto) — applies evolutionary algorithms to prompts and skill files. Reflects on reasoning traces, proposes "mutations," then uses Pareto-based selection to maintain a diverse frontier of high-performing variants. Compatible with DSPy.

• EvoSkill — applies GEPA's evolutionary approach specifically to multi-agent coding workflows. Tracks multiple skill candidates on separate Git branches; Pareto frontier selects winners on a held-out dataset.

Key trade-off flagged: all three require a verifiable feedback signal and a clean evaluation dataset — they cannot optimize on open-ended subjective tasks. Upfront compute cost is high (LLM optimizer reads voluminous trajectory histories), but inference costs are unchanged since the output is a plain text file.

"Loop engineering" is named as the broader architectural shift: assembling control systems with evaluation metrics, memory, and exit conditions rather than hand-tweaking prompts.

Sponsor item (self-cross-promo / paid): "From Tiger Data" — TimescaleDB/CERN webinar. Labeled as sponsor content; not editorial.

Mapping against Ray Data Co

RDCO runs a live skill harness at ~/.claude/skills/ with ~40+ skill files managed manually. The /improve skill already implements a rudimentary version of this loop (reads feedback, proposes rewrites), but it is human-gated and single-shot rather than automated.

Immediate relevance:

1. SkillOpt's "bounded edits" concept maps directly to the risk RDCO already experiences: fixing one skill breaks another. The textual learning-rate budget and rejected-edit buffer are concrete mechanisms worth pulling into /improve or a future skill-optimizer pipeline.

2. Held-out evaluation dataset requirement is the blocking prerequisite RDCO does not yet have. Each skill would need a task suite with verifiable pass/fail signals before automated optimization becomes viable. This is the actionable gap.

3. GEPA + EvoSkill's Pareto frontier with Git branches is the multi-hypothesis approach — directly applicable to skills where desired behavior varies across task types (e.g., deep-research needing different behavior for papers vs. market analysis).

4. "Loop engineering" framing reinforces the RDCO architectural direction: the COO agent's autonomous loops are already the substrate; skill self-optimization is the next layer to add.

Practical near-term path: the /improve skill could be upgraded to (a) log skill edits with before/after snapshots, (b) capture structured pass/fail signals from skill runs, and (c) use that history to propose bounded diffs rather than full rewrites.

Related

• [[~/rdco-vault/06-reference/2026-04-15-thariq-claude-code-session-management-1m-context]] — session management discipline that informs how optimization loops should handle trajectory context
• [[~/rdco-vault/02-sops/2026-05-18-implementation-notes-pattern-for-sub-agent-dispatches]] — sub-agent dispatch pattern; self-optimizing skill pipeline would extend this
• [[~/rdco-vault/02-sops/2026-05-19-verification-as-independent-worker-pattern]] — independent verification gate; analogous to SkillOpt's held-out validation set