| name | self-evolving-systems |
| description | Runs MetaHarness propose-evaluate-log-iterate loops to permanently improve agentic systems (workflows, prompts, retrieval logic, memory, orchestration). A coding-agent proposer inspects prior code + execution traces, diagnoses failure modes, proposes targeted edits or full rewrites, and iterates until Pareto-optimal — Karpathy's autoresearch pattern applied to the harness layer (not self-correction, but permanent self-evolution). Use when a workflow has been manually tuned but plateaued, quality-gate scores consistently land in the 6-7 range, the same class of error recurs despite self-annealing, the user wants better approaches than hand-tuned prompts, before upgrading to a more expensive model (evolve the harness first), or running periodic optimization sprints. Trigger proactively whenever the user says "this skill is stuck", "evolve the prompt", or "make this workflow better over time". For the business-harness deployment lifecycle and readiness audits use nate-b-jones-auto-improvement-loops instead. |
| expert | Composite (Yoonho Lee / Andrej Karpathy / Rich Sutton) |
| domain | AI/Automation — Self-Improving Systems |
| version | 1 |
| source | "MetaHarness — End-to-End Optimization of Model Harnesses" (arXiv:2603.28052v1, Stanford/MIT/Crafted) + "autoresearch" (Karpathy, 2026) |
Self-Evolving Systems
The harness around the model matters as much as the model itself. A 6x performance gap exists between the best and worst harness for the same weights. Hand-engineering harnesses is the status quo. It shouldn't be.
Core Capability
Design, run, and audit self-improvement loops for any agentic system component — workflows, prompts, retrieval logic, memory management, orchestration patterns. Uses the MetaHarness methodology: a coding-agent proposer that inspects prior code + execution traces, diagnoses failure modes, proposes targeted edits or full rewrites, and iterates until the system converges on Pareto-optimal performance.
This is not self-correction (recovering from errors). This is self-evolution (permanently improving the system).
What This Skill Produces
| Workflow | Output |
|---|
| Self-Evolve | Full MetaHarness-style evolution loop on any Antigravity component |
| Harness Evolve | Focused evolution of a specific workflow's prompts, flow, and gates |
| Auto-Experiment | Karpathy-style experiment runner with git-branch isolation |
| Evolution Audit | History inspection, Pareto frontier, regression detection |
| Skill Anneal | Self-annealing applied to skill prompts/workflows via past failures |
| Proposer Sprint | Time-boxed, fixed-iteration improvement sprint |
| Evolution Status | Dashboard of all active/completed evolution loops |
| Bitter Lesson Check | Audit hand-coded heuristics for evolution replacement potential |
| Trajectory Ratchet | Multi-cycle evolution protocol with monotonic progress enforcement — prevents drift, local optima, and coherence loss across hundreds of changes |
When to Deploy
- A workflow/skill has been manually tuned but performance plateaus
- Quality gate scores consistently land in the 6-7 range (good but not great)
- The same class of error keeps recurring despite self-annealing fixes
- You want to discover better approaches than hand-tuned prompts
- Before upgrading to a more expensive model (evolve the harness first)
- Periodic optimization sprints on high-value workflows
Stacking Opportunities
- + Nick Saraev (Agentic Workflows): Self-annealing + evolution = self-correcting AND self-improving
