White Bay

Field notes from a one-person shop

Nine hours of work
happen while you sleep.

A working guide to running coding agents unattended overnight — the setup, the isolation, the approval gate, and the morning review that tells you whether to trust the output.

run 1fb60f0e · isolated worktree 22 Aug · 23:41 → 05:02
23:41goal accepted — 13 tasks planned
00:07t01 scaffold committed
01:22t06 approval requested → held until morning
02:48t09 tests 17 passed
04:16t13 build 10 pages
05:02run idle — nothing merged yet
08:30you review, then merge
The agent never merges on its own. That rule is the whole method.

What this actually is

Not a course about the future of AI. A record of one setup that runs every night, written down completely enough that you can rebuild it.

Every command, in order

Install through first successful overnight run. Including the failures — the linker error, the terminal that hangs on a cursor query, the process that won't die.

Isolation before autonomy

Each run gets its own worktree. Nothing touches your main branch until you say so. How that's wired, and why it's non-negotiable.

An approval gate that holds

The agent asks before it spends, sends, or publishes. Approvals arrive on your phone. Approving twice does the thing once.

Morning review, not vibes

How to tell a finished run from a stalled one when the board lies to you. Commit-first verification instead of trusting the status.

Contents

Written in the order you'll do it. Each chapter ends with a state you can verify before moving on.

  1. The machine you needToolchain, build dependencies, and the two installs that cause most first-day failures.
  2. First run, watchedA small goal, run start to finish while you sit there. What normal looks like.
  3. IsolationWorktrees, branch policy, and cleaning up runs that ended badly.
  4. The approval gateWhich actions require a human, how the prompt reaches your phone, and idempotency.
  5. OvernightQueue design, what to leave running, what never to leave running.
  6. Morning reviewVerifying by commit and test output. Deciding merge, rerun, or discard.
  7. Shipping the resultStatic deploy, checkout integration, license keys, and verifying a payment webhook.

Pinned versions

Guides like this rot because versions move. Every command in the book is run against exactly one set, recorded here before release. When they change, buyers get the revision.

ComponentVerified againstStatus
Agent runtimeOrca 1.4.187verified
Claude Code2.1.229verified
Rust toolchainrustc 1.97.1verified
Node24.19.0verified
Host OSWindows 10 · build 19045verified

Get the guide

One payment. PDF plus the working files — prompts, settings, and the scripts referenced in each chapter. Revisions are included as versions move.

$39one time · instant download

Roughly 120 pages, plus a file bundle you can drop straight into a project. Written for people who already write code and want the unattended part to work.

Buy — $39 Read a sample chapter first

Before you buy, read the contents above and the free sample chapter — they show you exactly what is inside. If something breaks, email support and I will fix it; if I can't fix it within seven days, you get a full refund. Full policy.