Autonomous Claude Ops
yurukusa/autonomous-claude-ops
Audita atividade do Claude Code e detecta movimento sem resultado
Ver no GitHubAgent que audita logs de sessões do Claude Code para identificar o padrão "máquina ocupada" — quando há muita atividade (centenas de respostas) mas pouca saída concreta. Fornece ferramentas e metodologia para diagnosticar e corrigir falta de progresso real em operações autônomas com IA.
Instalação
npx convez-ai yurukusa/autonomous-claude-ops1 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
- Diagnosticar por que agentes IA geram muita saída mas pouco resultado tangível
- Analisar padrões de sessão para identificar desperdício de tokens
- Implementar guardrails para evitar trabalho repetitivo ou inútil
Veredito

Ferramenta prática que aborda um problema real e pouco discutido: agentes IA que trabalham muito mas entregam pouco. O script `verify.py` é simples e direto, útil para quem quer medir o próprio viés de atividade versus resultado.
README
A tradução para o português deste README ainda está na fila.
Autonomous Claude Ops
Your AI agent isn't lazy. It's busy at the floor — and that's worse.
This is the free companion to the short book Autonomous Claude Ops. It contains a tool you can run today, a free sample chapter, and the core idea, so you can check the claim against your own machine before reading another word.
The 60-second version
I ran Claude Code autonomously for a month and built every safeguard against the obvious fear — that it would sit idle. It worked: the machine almost never went idle. Then I counted my own session logs:
- 295 sessions, a median of 488 assistant responses each
- 85% of sessions did 100+ responses — the machine was a firehose
- only 2% did nothing
…and in the same month it produced 47 product folders, 36 finished-but-unpublished drafts, and almost nothing that reached a single reader.
The failure mode of autonomous AI operation is not idleness. It's motion that never reaches anyone, and a tireless agent generates that comforting motion faster than a lazy one. This repo and the book are about catching it.
Run the audit on your own logs
verify.py reads your local Claude Code session logs (read-only — it writes nothing) and shows the shape of your own activity:
python3 verify.py
It reports your session count, the median responses per session, and how many ran past 100. If the median is in the hundreds and your outward output (things published, shipped, sold) is near zero, you are where I was. (For a richer version that also breaks down where your tokens actually go, see cc-ops-audit.)
Free sample chapter
Chapter 1 — The Busy-Machine Trap is included here in full. It's the diagnosis; the book is the fix.
What the book covers
Six short chapters, one idea aimed at six surfaces, every claim checkable against your own logs — deliberately not another 500-page "build powerful agents" manual:
- The Busy-Machine Trap — why activity isn't progress (free, in this repo)
- The Outcome Ledger — redefine progress so busywork can't count as a win
- The Gate Before the Action — two questions that kill floor-pointed work before it starts
- The Silent Failure — when a tool reports success but changed nothing (a confirmed
ENOSPCcase), and how to make the silence loud - The Drain You Can't See — where your tokens really go (it isn't the size of the edit)
- Surviving the Session Boundary — the one handoff rule that stops every restart from re-deriving what the last session settled
Get the book
Autonomous Claude Ops is coming to Kindle (USD 9.99). Star/watch this repo to catch the release — the buy link lands here when it's live.
License
The tools in this repo (verify.py) are MIT licensed — use them however you like. The sample chapter is shared free; the full book text is not under the MIT license.