self-evolve
Stealthy-McStealth/self-evolve
Agentes Claude Code que aprendem com seus erros e mantêm consistência
Ver no GitHubPlugin que transforma um agente Claude Code em um sistema autoaprendizado persistente, similar ao Hermes. Extrai habilidades duráveis de execuções anteriores, as armazena e impõe conformidade em trajetórias longas para evitar que o agente desvie de suas instruções ou repita erros.
Instalação
npx convez-ai Stealthy-McStealth/self-evolve1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Desenvolvimento iterativo com aprendizado multi-sessão
- Automação de tarefas longas mantendo foco e instruções
- Construção de bibliotecas de skills reutilizáveis no Code agent
Veredito

Implementação rigorosa (com fundamentação em pesquisa peer-reviewed) que resolve um problema real: Claude Code sem persistência de aprendizado. A abordagem de extração de skills e enforcement de aderência é sofisticada, embora exija que o usuário domine o sistema de plugins do Claude para extrair o máximo valor.
README
A tradução para o português deste README ainda está na fila.
self-evolve
Make your Claude Code agent work like a real harness — the one Hermes uses.
an agent that can't learn from its mistakes isn't an agent. it's a very expensive grep.
A Claude Code plugin that gives your agent self-evolving harness capabilities: it learns from mistakes, encodes durable skills, and actually follows them across long trajectories. The same architecture that makes Hermes effective — applied to your local Claude Code setup.
Backed by peer-reviewed research. Not vibes.
why this exists
Hermes-style agents don't just run tools. They build up a persistent skill library from execution traces and stay on-task across long trajectories. That's what makes them effective over time — they compound.
Out of the box, Claude Code doesn't do this. Every session starts fresh. Skills drift mid-conversation. You repeat yourself.
This plugin fixes both problems with two mechanisms:
- Skill evolution — extract learnings from conversations and persist them as reusable skills (like Hermes's harness updating)
- Adherence enforcement — prevent mid-trajectory drift where the agent loads instructions but quietly stops following them
the research
This plugin implements findings from:
"Harness Updating is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents" Lin et al., arXiv:2605.30621, May 2026
Key findings:
| Finding | Implication |
|---|---|
| Any model can write good skills (spread: 3.1pp max) | Skill-writing is pattern extraction, not deep reasoning |
| Performance is bottlenecked by the agent, not the evolver | Invest in following skills, not writing them |
| Weak agents fail to activate skills (25% vs 96%) | Trigger phrases and explicit invocation matter |
| Adherence decays 4x faster in weak models over long trajectories | Always-on drift prevention is essential |
what's included
self-evolve/
├── CLAUDE.md # always-on adherence protocol
├── skills/
│ ├── improve-urself/SKILL.md # /improve-urself — encode learnings as skills
│ └── reload-context/ # /reload-context — fix mid-trajectory drift
│ ├── SKILL.md
│ └── references/
│ └── skill-standards.md # template for writing followable skills
└── .claude-plugin/plugin.json
skills
/improve-urself [optional: topics]
Scans the current conversation for mistakes, corrections, and patterns. Creates or updates a persistent skill so the learning survives beyond this session.
/improve-urself
/improve-urself bash error handling
/improve-urself json formatting tool-use
The key insight: you don't need a frontier model to write the skill. The quality is bounded by what the execution trace reveals — not by how smart the writer is. Even small learnings compound across sessions.
/reload-context
Re-reads every loaded skill and active instruction. Explicitly re-commits to fallback steps. Makes the active harness visible and auditable.
Use it when:
- Claude is improvising around instructions it should be following
- A task is getting long and you want to reset attention
- Output looks right-shaped but something feels off
CLAUDE.md — always-on adherence protocol
Runs every session without explicit invocation:
- State relevant loaded skills before multi-step tasks
- Re-check alignment at sub-task transitions
- Acknowledge fallbacks upfront, not after step 1 fails
- If improvising around a skill → stop and re-read
install
Claude Code CLI
claude plugin add Stealthy-McStealth/self-evolve
Manual
Clone this repo into your Claude Code plugins directory:
git clone https://github.com/Stealthy-McStealth/self-evolve.git ~/.claude/plugins/self-evolve
limitations
- Non-monotonic benefit: mid-tier models benefit most from evolved skills. Very weak models struggle even with perfect skills because they can't reliably invoke or follow them.
- Not fully automatic: you call
/improve-urselfyourself. Fully automatic skill evolution from conversation traces requires more infrastructure — that's future work.
contributing
PRs welcome. If you've found a pattern that makes skills more followable or adherence more durable, open an issue or submit a skill.
related
- Paper: Harness Updating is Not Harness Benefit — the research this implements
- A-EVO-Lab/a-evolve — code from the paper