Bounded Autonomy Kit — run your AI agent without getting burned
Your AI agent will eventually do the unrecoverable thing.Give a coding agent enough rope to work unsupervised and you hit two opposite walls: too tight (you approve every step, so you're the bottleneck) or too loose (it sends the email, ships copy with a client's real name in it, deploys the regression, or force-pushes over your work — and now you're cleaning up).The Bounded Autonomy Kit is the middle path: the agent moves at full speed on everything reversible, and stops at a human gate only for the handful of actions that can't be taken back. It's the exact set of documents and the one script the author runs every day as a solo operator with an AI doing most of the work.What you get (8 parts, ~15 files) • OPERATING_MODEL/ — the doctrine: autonomy by default, the exact reversible-vs- irreversible line, and the operator/agent handoff shape. • GATES/ — a zero-dependency Node script (ship-gate.mjs) that scans any file, folder, or piped text for client names, secrets, private paths, and leaks before you publish or send, and BLOCKs on a hit. Plus the editable rule file and hook wiring so it fires automatically. • PREMORTEM/ — the death-before-the-fact check that fires *precisely when you're most confident*, with worked examples. • DAILY_OPS/ — a nightly self-review and a weekly one-pager that keep the agent shipping instead of polishing forever. • JUDGMENT/ — a three-lens council prompt (builder / skeptic / recipient) for the genuinely hard calls, with a filled-in session. • INCIDENTS/ — six real postmortems from running an AI agent solo, each ending in a concrete rule so the same mistake can't recur. • EXAMPLES/ — a ready-to-paste agent instruction block, grounded in one fictional team so nothing exposes real data.Why it's built this way • The line is reversibility, not importance — so the agent never has to guess how much something matters, only whether it can be undone. • A machine gate, not eyes — because the leak that burns you is the one a tired human slides right over at 1am. • Premortem on confidence — because an agent's confidence is not evidence. • Incidents become gate rules — so lessons move out of your head and into a script.Setup in about 10 minutes. Node 18+ for the one script; everything else is Markdown. Works with Claude Code and any agent that reads instruction files in your repo. The gate is plain Node you can call from a hook or CI step.Who it's for • Developers running an AI coding agent unsupervised who've felt the "too tight / too loose" problem • Solo builders letting an agent do outbound, deploys, or publishing on their behalf • Anyone building toward an "AI-run solo company" who needs the agent bounded, not cagedBuy once, use in every project you own.
Get it → plainbuilt.gumroad.com