More detail: Brief your Chief of Staff to build a third team while you're away, and check it when you're back
Optional reading for this lesson. Everything you must do or decide is in the lesson.
This page goes with the lesson (course/workforce/week3/lessons/anw-w3-l10.md). You don't need it to finish the lesson.
What your Chief checks before it plans
Three things the run can't fix for itself, because they live in files the run may never touch. Your Chief checks them before its questions and tells you what is missing; the plan refuses with one sentence naming the fix while any is missing:
| Check | READY when | The fix |
|---|---|---|
| Your material is filed and approved | Your approved sources list at least one file for the team's department (or for your whole organization), each file exists, and each library file has a source card in your index. | Name the files from your library folder. The Knowledge Librarian (Sallee) files each as a source card and approves it for the department as part of your yes to the plan; there is no separate yes. |
| Hatch's source map names your work | work/course/staff/hatch-source-map.md names your work, not the course's example, with one place Hatch may read, or your word that the team reads files only. | Ask your Chief for the tools interview again, or say the team reads files only. |
| The before printout is saved | The Charter Steward's registry printout is saved in the job's folder and names today's agents. | Ask your Chief to have the Steward print your registry. |
What your Chief saves before the start
After your answers, your Chief of Staff saves six short files in the new job's folder, runs/<date>-overnight-<name>/overnight/contract/. Together they are everything the run may start from:
| File | What it holds |
|---|---|
brief.md | Your words first, word for word, then your Chief's version: the outcome, who receives it, your starting point, the department, the five roles and each agent's kind, what the team must never do, and the assumptions you allowed. |
acceptance.md | Four of your five test cases (case 1, case 5 and two of cases 2 to 4), each with Input, Should and Fails if lines. |
heldout.md | The sealed case: the remaining one of cases 2 to 4. |
allowlist.md | The files the agents may read during the run: your approved sources for the team's department and the ones for your whole organization, plus your founder brief and priorities. Never a mailbox. |
tools.json | The checking scripts the run may use, by name. No connected tools while you're away. |
caps.json | The limits: the most hours (6 by default), a latest finish time (6 AM by default), minutes for each step, at most two repairs and two rounds of questions to Archie. |
Material from a tool you've connected comes in as a file. Save a copy into your library folder and add it to your approved sources before the plan, the same way as for your second team.
The plan id, and why your Chief carries it
The plan prints a fingerprint of those six files, the plan id, and the longest the run can take at worst, from the limits in caps.json, so you can see it finishes before the latest finish time. When you say yes, your Chief records it against that id ("plan <id>, approved on the owner's yes: '<your words>'"), so your yes covers exactly that plan. You never type the id. If any of the six files changes after you've seen the plan, the start refuses: your Chief shows you a new plan, with a new id, and asks again.
The pre-authorization line
At your yes, your Chief offers one line, once: "If every case passes, install it; only stop me for a Needs-you item." Say it in your own words only if you want the install to go ahead when you're back without a second yes. It is saved with the date under your brief. Gigawatt honors it only when it can read, from the files, that every new agent's render check passed, no blocking finding is open, every visible case and the sealed case have their rehearsal receipts, nothing waits on your Needs-you list, and the run's own decision was "install as staged". Any one missing, and the install waits for your yes, with the reason named. A plain "yes" is never read as a pre-authorization.
What the dry run checks
The dry run starts one short session exactly the way the real run will start: the same app program, the same settings, the same folder. It checks three things: every checking script the run needs can be reached, a write outside the job's folder is refused, and a helper can be started. It also grades one agent already on your record with the same check the new agents get: a good agent should pass, which shows the check is fair before anything new is judged. It takes under a minute in Claude Code and a minute or two in Codex. The run starts only after a dry run of the same plan has passed.
What runs while you're away
Each step is its own fresh session of your app, run in the background by the Python file your Chief started, workforce/house/overnight.py. You can start it at any time of day; overnight is just one good time. Each step has one lead:
- Design. Archie writes the design, a numbered PRD such as
prd-v1.md, with helpers that read your sources and check the draft. Where it would ask you, it takes the safest reading and marks it as an assumption for when you're back. - Build. Gigawatt prepares each new agent in
runs/<job>/setup/, for both apps, and writes the proposed team record. Each new agent also gets a research file, a scorecard from two reviewers, a version 1.1 kept only if it scores higher, and deployment notes. If it needs a fact only Archie can give, it asks Archie in writing, at most twice. - Check. An independent session reviews each prepared agent on its own. An agent that fails goes back to Gigawatt to fix, at most twice.
- Rehearse. Your Chief's background session runs your test cases on the prepared agents: case 1, then the sealed case, then case 5, then the others. Every case writes a receipt that says it was a staged rehearsal.
While you're away nothing is installed, recorded, sent, published or connected. Before the first step, the run takes a fingerprint of every file an install or a record would change: your agent folders, skills, profiles, registry, records, settings, charter and library. It checks them after every step. Any change stops the run and is reported as VIOLATION.
The sealed case
Your Chief keeps one of cases 2 to 4 apart from everyone building the team. Its words never appear in any other file the run reads; the plan refuses if they do. It is shown only to the last step, so your team gets one fair test it wasn't built on. Its facts may still be in your material, the way a real job's would be. In Codex the seal is kept by instruction and checked afterward; the report says whether it held.
How the run ends
The status you read when you're back gives one of these:
| Status | What it means | What to do |
|---|---|---|
done | Every step finished, and every case has a receipt. Some agents may still be marked needs you. | Read the report, starting with Needs you, then observe the cases. |
incomplete | A step ended without everything it should have left. The report lists what wasn't produced, and anything only you can change under Needs you. | Fix what's yours, then ask your Chief what finished, and to carry on from there on your new yes. |
stopped | You stopped it, or your off switch did. | Ask your Chief what finished. |
exhausted | It reached the latest finish time or a step's time limit. | Ask your Chief to carry on from there, on your new yes. |
paused-usage | Your app's usage limit was reached. | Carry on later, on your new yes, once your limit resets. |
blocked | A fix for the same finding kept failing, or a decision only you can make blocks everything. runs/<job>/overnight/BLOCKED.md says what. | Read it with your Chief, decide, then carry on on your new yes. |
violation | A protected file changed while you were away. | Don't install. This job can't carry on: look at what changed with your Chief, and start a new one if you want. |
Carrying on reruns only the steps whose files changed or never finished. Finished work is kept.
Stopping it
Say "Stop the background build." in your Chief's thread. The off switch from your email department works here too: the same autonomy-off file in work-orders/standing/ stops a running background build, and a new one won't start while the file is there.
The report
runs/<job>/overnight/morning-report.md (the file keeps that name whatever time you read it) is written from the files, not from anyone's memory. In order:
- The status line and the protected-files line.
- Needs you: anything only you can change: a source that isn't filed or approved, a rule, a borrowed agent's limits, or a case that asks an agent to do what it never does. Each names the file it lives in. The run never tries to fix these in the background; they wait for you. A first build often ends with one or two.
- The agents: for each new one, whether it was built, its check verdict, whether it was staged and rehearsed, its scorecard before and after version 1.1 (a note, never a gate), and its status:
ready for your yes,needs you, or not ready. Concerns the check noted but didn't stop on are listed under the agent; they ride along into your decision and don't stop an install. Onlyready for your yesagents go into the install. - The cases: for each, whether it ran, whether it has a receipt, and whether the receipt says it was a staged rehearsal.
- Anything refused during the run. A required step that was refused makes that step incomplete.
- Whether the sealed case stayed sealed.
- The decisions and assumptions waiting for you, from the design's interview notes.
- The limits, and how much of each was used, ending with one line to read:
This run used about N tokens (about $X). If that is more than you expected, lower the limits in the plan next time; your Chief sets them incaps.jsonfrom your words. - Below a fixed line, the run notes, written by a model: a short story of the run, the blanks it left for you, and one suggestion: install as staged, send listed changes back to Archie, or drop it. The suggestion is a note. The decision is yours, unless you pre-authorized the install: then "install as staged" is one of the conditions that let it go ahead.
Two doors, and what rides on the second
The design yes is the first door: before you go, to the brief, the cases and the plan. The install is the second: when you're back, in Gigawatt's thread, one yes in your own words, and Gigawatt records both the design version it installs from and the install number; the install refuses if a staged file changed after your yes. The team record is filed by your Chief with that install, your yes written in, and because this is your third team, the same filing turns the Charter Steward's daily check on. Neither needs a yes of its own. While you were away, your yes before the start covered only the design, the preparing, the checks and the rehearsal. Between your two yeses the team stops only for five things: money, sending anything outside, deleting, posting publicly, and account authority. Each one waits on your Needs-you list with the work saved up to that point; everything else keeps going.
In each app
| Claude Code | Codex | |
|---|---|---|
| Before the start | The app asks to run the start. Allow it. | The app asks to run the dry run and the start outside its sandbox. Approve both. |
| While it runs | Each step's writes are limited to the job's folder by the app's own sandbox and settings. | Each step runs locked to the job's folder, with your connected apps turned off for the run. |
| What the report can show | Every agent's actions in each step. | The lead's actions in each step. Helpers report in the files they write. |
| How long | Plan on a few hours. | Plan on longer than Claude Code. |
On Windows
The background build isn't available on Windows yet, and nothing keeps a Windows computer awake. Your Chief says so and offers to build the same team with you in the daytime, the way you built your second team.