AI Build Lab / Learn · Agent Workforce · Week 3 · Module 02 · Details

More detail: Plan your email team's schedule, and how you can entrust your team to do more without you

Optional reading for this lesson. Everything you must do or decide is in the lesson.

Week 3 · Module 02 · Lesson 3 · Details

This page goes with the lesson "Plan your email team's schedule, and how you can entrust your team to do more without you" (course/workforce/week3/lessons/anw-w3-l04.md). You do not need it to finish the lesson. Read a section when the lesson links to it.

How the Chief works out the interval

The request is arrival times over the last two weeks. Codex Outlook may return subject and message text with a list; confirm that reading scope first, or stop without searching. Count only the evidence actually returned, and distinguish Outlook messages from real conversations. From the returned times, the Chief reads how many listed items arrive a day and which day was busiest. From the record, it reads how long the team took per email in your three observational evals. It suggests the most frequent interval the team can still finish on your busiest day.

The Chief needs at least three timed emails in the record. If it has fewer, run another check of your own inbox. Nobody types a number in to get past this.

Time zones. If your working hours are in one time zone and your computer's clock is in another, the Chief says so. The schedule's own times can then look different, but the times the Chief reads out in your working hours are what you get.

What a schedule does, in full

Missed times. A schedule runs only while your computer is on and the app is open. A time the computer sleeps through is skipped. In Claude Code, the app may start one catch-up check for the most recent missed time when the computer wakes, so a morning check can run later in the day. In Codex, plan as if a missed time is simply skipped.

Never sends. A scheduled check reads mail, sorts it and updates the record. In Codex Outlook Email, the Chief never sends at any time. The student may manually send an eligible reviewed draft only when no hold, full stop or off switch is active. Other routes where the Chief can send keep their own version-bound approval.

Rung 1 on a schedule. On day one every category is on rung 1. So a scheduled check sorts every new email, lists the rung 1 ones under Needs you as "waiting for a manual run", and drafts nothing for them until you run a check yourself. Your own "check my email" handles every rung.

Overlap. If two checks would overlap, one says "still busy" and stops, and the record shows it.

Picking up again. If the app quits, or its usage limit stops a check partway, the record has every step up to that point, and the next check picks up from there. Mail seen but not yet sorted comes first. A change to your mailbox that was started and not confirmed is settled by checking the email thread, never by doing it again.

When the schedule switches on

The "What the course has tested in your app" table in the previous lesson says when creating the scheduled task opens for your app. Then the Chief guides you through creating it yourself, one step at a time, and the record has to show one scheduled check before anyone calls it working.

Why not the cloud. A cloud run would work from a copy of your project on GitHub, and your record, drafts and mail are kept out of GitHub on purpose. None is set up this week.

Moving up and down, in full

Codex Outlook boundary. A read-back draft or a student's report of a manual send is not a confirmed send or an Unedited/Light result. Keep manual reports and missing evidence separate from measured outcomes. Do not promote a category or infer autonomy credit from them. The following outcome/promotion rules require the independent evidence they name; this route does not manufacture that evidence.

Your copy. Read work/course/staff/autonomy-ladder.md beside course/workforce/week3/autonomy-ladder.md. The normal ladder check creates a missing copy or refreshes one you have never edited; it keeps a customized copy unchanged. Your customizations remain yours, within the current runtime limits: no older copy overrides Codex Outlook's drafts-only restriction, holds, practice safeguards, full stops or the off switch. A manual-send report is not confirmed-send evidence or promotion credit. If a customized copy lacks the Outlook note, the Chief shows a proposed additive Outlook note, preserving every existing customization. It waits for your explicit yes before writing that exact addition, never silently replaces a customized copy and never claims it is saved before a readback. This is consent for editing a customized file, not a new approval for the normal unedited-copy refresh. The restrictions already apply while you decide.

Edits in your mail app. If you change a draft in your mail app before you tell the Chief to send it, the change is placed as a new version, with a fresh review, and you say yes to that version.

Why "light" is so strict. The record compares the sent copy with the draft the team placed, and anything it cannot tell counts as Rewritten. Changing $150 to $175 costs the category a rung and costs you nothing else. A changed price counted as "good" would cost you much more.

The other counts. Per category, the team also counts the typical time from draft to send, false skips (emails it skipped that you then said mattered), and drafts the Quality Reviewer (Kansa) still had concerns about after one revision.

Moving up needs two things, then your yes.

  1. Twenty clean drafts. The last 20 finished drafts in that category were all Unedited or Light, and you reviewed every one. Practice emails never count.
  2. A replay that passes. Kansa checks drafts for five older emails in that category that the team has not seen. Every one of Kansa's twelve checks must pass on every draft. A failed replay blocks the promotion, and your yes does not override it.

When both hold, the Chief shows you the evidence and asks. The Charter Steward files your yes as a decision record first, with your exact words. Only then does the Chief record the new rung. After you change a category, a rule the drafts rely on, or the team's models, the count starts again for the categories that change touches.

Moving down. A category drops one rung the same day, without asking you, when you send a Rewritten draft, discard one, or Kansa still has concerns after one revision. Rung 1 stays at 1. Moving back up needs the evidence above.

The off switch and every full stop

The off switch pauses team placement and sending. It cannot disable Outlook's Send button. While a hold, full stop or off switch is active, the student must leave team drafts unsent too. Clearing a stop resumes eligible Outlook drafting, never agent sending.

The off switch is a file named autonomy-off in work-orders/standing inside your workbench. Name it exactly that when you can. Any file whose name starts with autonomy-off counts, so autonomy-off.txt, which Windows may add when file name extensions are hidden, still stops sending. While it is there, nothing sends and no new draft is placed in your mailbox. The Chief never creates it, because that folder is yours alone. Use it with your app's own stop and the prompt before each send.

Every full stop. These stop all sending, for every category, until you look and say go:

  • a draft or send to the wrong person or thread
  • a send you did not approve
  • the same email sent twice
  • an agent acting on instructions found in an email or attachment
  • details from one person's email turning up in another's draft or in a web search
  • a draft touching a password, bank details or a payment
  • the team acting on a draft it did not place as if it were its own
  • a job writing to an email after it has let go of that email
  • your off switch

A draft, recipient or calendar fact that changed after you approved it is caught before sending, and you see the new version. Only your words clear a stop.

What The Professor does

The Professor explains, defines, says where you are in the course, checks your saved work against a lesson (met, not yet, or cannot tell), and helps when a setup step fails: it asks what you saw on screen, reads only what the lesson allows, and writes up what failed, what it checked, what you can try, and who to ask next. Anything that must be done goes to your Chief of Staff.