rrsi vs SkillOpt
Google Research's Regularized Recursive Self-Improvement: evolves an agent harness (prompts, tools, memory, sub-agents) with constraints on the search so gains hold out of distribution. — versus — Microsoft's text-space optimizer that trains a frozen agent's skill document like weights — rollouts, bounded edits, validation-gated updates — and ships a compact best_skill.md.
Same discipline — accept an edit only when held-out evaluation measurably improves; SkillOpt optimizes one skill document, RRSI the whole harness (prompts, control flow, tools, memory).
| rrsi | SkillOpt | |
|---|---|---|
| Stars | 1.5k | 18k |
| Forks | 141 | 1.7k |
| Language | Python | Python |
| License | Apache-2.0 | MIT |
| Last activity | 18 days ago | 5 days ago |
| Topics | agents, training | skills, training |
| Curated connections | 3 | 6 |
rrsi — the curator's take
Read and run it if you evolve harnesses against a benchmark and keep watching in-distribution gains evaporate — annealed edit budgets, a leakage critic, a noise-adjusted acceptance floor and a cost rule are the fix, each mapped from paper to code, with every candidate a git worktree. Not a product or a plug-in: it's paper code wired to Claude Opus on Vertex AI and three specific benchmarks (Terminal-Bench, Harvey LAB, EngDesign), and a run costs real compute.
SkillOpt — the curator's take
Use it when you have a task with a scorer and want a skill that measurably improves on it — an edit lands only if the held-out score rises, and deployment adds zero model calls. Not a passive learner: you need data, a benchmark and an optimizer-model budget; the nightly SkillOpt-Sleep mode is the bridge to everyday Claude Code/Codex sessions. No metric? An in-use skill learner fits better.