Reference: self-improving harness + self-evolving skills

Owner intake 2026-07-08. Two sources, digested for the nexus stack. This is the REFERENCE copy (toolnexus); the build SPEC lives in the brain repo (brain/docs/SPEC-self-evolving.md).

Where things live (the decision)

Concern	Home	Why
Learning lifecycle: experiences → reflect → curate → evolving playbook → regression-gate	brain	brain is already the persistent, git-backed, evidence-first store with record/recall/consolidate/check; the new verbs extend it
Skill library: versioned skills, failure-driven synthesis, validate/deprecate, metrics	brain	same store, same git genealogy; skills are learned artifacts
RUNTIME use of evolved skills by agents	toolnexus	toolnexus builds/runs agents; it LOADS brain playbook + brain skill search output as agent context/tools
OS doctrine (who runs the loop, when)	deemwar-one-os doctrine/self-improving-harness.md	cadence + guardrails for CEO/lines/desks

One-liner: brain learns, toolnexus acts. No new repo needed.

Source 1 — Lilian Weng, "Agent Harness" (lilianweng.github.io, 2026-07-04)

A harness = everything around the model: plan/act/observe loops, tools, memory, evaluation, permissions. Key adoptables:

• Evolving playbook, not a growing text blob: structured itemized entries; a reflector distills each trajectory (success AND failure) into deltas; a curator merges/dedups/ supersedes.
• Regression-gate self-edits: any proposed harness/playbook change is checked against held-in cases that already worked; rejected candidates are logged, not applied.
• Failure root-cause, verifier-grounded — record the pattern + cause, not the surface error.
• Guards: evaluator OUTSIDE the loop (no self-grading), anti-diversity-collapse (probe poor-looking paths on paper), negative results first-class, human up the stack (strategy + irreversible forks only).
• Prereq: recursive improvement only works with a capable base model.

Source 2 — arXiv 2605.23904, "SkillOpt: self-evolving skills"

Closed-loop skill lifecycle: discover → verify → curate → reuse → refine.

• Skill = executable procedure with id, description, preconditions, success criteria, cost (calls/latency), metrics (success rate, invocation count).
• Failure-driven synthesis: repeated failure patterns generate candidate skills.
• Validation-gated acceptance: candidate must beat prior version on a held-out set (e.g., ≥5% improvement) or it's archived.
• Library: versioned repo, primitives + composites (chaining), genealogy (what replaced what → rollback), deprecation of low-usage/degraded skills.
• Explainability + human-in-loop: every change stores its rationale; manual curation coexists with automated.

Rollout in our OS

1. brainlib implements the new brain verbs (reflect, curate, playbook, regress, skill register|validate|search|metrics|deprecate).
2. toolnexus agents load brain playbook/skill search at spawn (runtime consumption).
3. Fleet contract: reflect+curate on every work done/block; desks at session end (crypto desk directive sent 2026-07-08); CEO retro = curator + regression pass.