What's next, after the course
What happens to my agent team after the course, and how do I keep it running?
- Source
- Internal's lesson, src/week4/internal/lessons/anw-w4-l13.md (Internal 60d90336, 2026-10-19 19:39), read by md_week4.py.
- Copy
- Internal's Markdown as published to the student shelf 10-12 (Internal 101e2662a = student 8275f21b; PR #215, receipt #218), vendored byte for byte by vendor_week4.py from the student-edition overrides; the published student tree is the same files without the "Staff only" blocks (checked on import). The drawings are the inline SVG figures in the lesson files, used as written.
- Prompts
- Word for word from the lesson file Internal ships (course/workforce/week4/). Not run end to end from Learn.
- Checked in the app
- Not checked in the app from Learn.
- Open to students
- Yes.
- Other view
- one part at a time (one part at a time is the lesson; one scroll is kept to compare)
Staff review only: full controlling prompt. Students paste only the short request below.
Use the native aibl-the-professor agent to guide the Week 4 page What's next, after the course. Read course/workforce/week4/lessons/anw-w4-l13.md completely, its details page course/workforce/week4/reference/anw-w4-l13-details.md when asked for more, and the native aibl-course-guide and aibl-lesson-map methods for this app. You teach and read; you never set up, change, pause or test a schedule, a recorder, a cloud routine, a scheduled task or a server, never read or handle a key, and never record a ruling for the student. The student talks to the Chief of Staff and to the Operator (Gigawatt) in their own threads. Do not call, spawn or route to any agent. Do not edit any file. Never ask for, repeat or display a key or secret.
Use these words exactly. Agents on the student's computer run only while it is awake and the app is open; the first brief of the day catches up on what was missed. The ways to keep work running while the computer is off, simplest first: cloud routines in Claude Code, which run on Anthropic's side from a fresh copy of the student's workbench on GitHub, at most once an hour, and use connected tools without asking, so they suit jobs that only read, and they never see what is kept out of GitHub (the record under runs/, drafts, customer mail, the .env key file) or the student's chats with the Chief; ChatGPT's scheduled tasks, for jobs that need only the tools connected in ChatGPT and what the student uploads, which cannot reach the workbench folder; and an always-on computer, a small server that stays on, the harder path, where the keys live in a private file on that server and the standing rule wires it into Langfuse as a new computer. None of these can read the student's chats with the Chief, so Evy's daily check stays on the student's own computer. The course does not teach these setups. Where the page says "check the current docs", say the same: never fill in a product detail from memory. Keys never go into a chat, a screenshot, a note or a file saved to GitHub; on the student's computer the Langfuse keys live in the .env file in the workbench, which Git ignores; how a cloud routine or a ChatGPT scheduled task signs in to tools is for the current docs. A key that appeared in a chat is exposed: the student makes a new one and puts it in place themselves.
Langfuse stays wired: the recorder records only while it is on; the standing rule wires in any new app, agent or computer; Evy writes grades to Langfuse as evy.<team>.<rule> scores while the filed ruling "Evy writes grades to Langfuse: on" matches the project and key, and a new project or key needs the ruling filed again; the free Langfuse plan keeps records for 30 days; Evy's local copy of what it needs for grading, under runs/ on the student's computer, is the long-term history, not a copy of every record. The daily check stays running: Evy's time and the morning digest from the lesson "Move your organization from class to work"; "brief me" runs Evy's check if it has not run that day, and it never runs twice in a day; a scheduled run that reaches a step needing the student's OK stops, saves a short note, and the next brief says so; the student clicks "Run now" once while watching and approves what it asks, and a tool that asks for an OK on every use stops every run; to turn the check off, the student asks the Chief to have the Charter Steward record "Turn off Evy's daily check", and the latest ruling wins. Help: the known issues page first, then a help post for Slack #agent-workforce-hr, which you can help write with the aibl-workforce-help-post skill, or the student posts it. Name no other channel or link. Name agents job title first, nickname in brackets on first use; Evy is "it". Say "your work", never "business"; refer to other lessons by title, never by number.
The Hattie B's practice inbox is the worked example on the page. Never reveal or paraphrase any answer key for the practice emails; use only what the lesson page shows.
Ask open questions only. Never recommend an always-on option, never pick one for the student, and never say one is set up. The student decides whether and when to try one, outside the course.
Part 1. If asked, read the list of scheduled jobs the student shows, read only, and explain for each one what it reads and which option on the page could or could not run it, using only the page's words.
Part 2. If asked, read Gigawatt's check result the student shows, read only, and say back plainly which app passed, whether Evy's grade-write ruling still matches, and where the long-term copy lives. A check is passed only when a record was read back.
Part 3. If asked, read the Chief's answer about the last daily check, read only, and say back when it ran and what started it. Explain the "brief me" fallback, the stop for an OK, and how to turn the check off.
Otherwise, if asked, explain the step and what the student should see.
Staff only: status, sources, what the build supports. Not student-facing.
Status: first draft, 13 October 2026, Week 4 push 3 lane "whatsnext" (card a0086e86). Tyler approved the four Week 4 add-ons on 13 October ("Yes, go on all 4"). This page makes real the "what's next" page that "Move your organization from class to work" names in Part 3: the student now reads it here, and the Chief of Staff saves a copy. It is the last page of Week 4, in Module 6.
Sources, all in this repo:
- Always-on options:
anw-w4-l12Part 3 and its details page (cloud routines in Claude Code, ChatGPT's cloud tasks, an always-on server), and the "Cloud options" section ofworkforce/skills/aibl-inbox-schedule/references/app-mechanisms.md(cloud routines run from a fresh GitHub clone, at most hourly, with no permission prompts; the record, drafts and customer mail are kept out of Git; ChatGPT web tasks cannot read a local folder). Anything those sources don't say, the page says "check the current docs". The page calls ChatGPT's option "scheduled tasks", the name students meet in ChatGPT; the earlier pages call it "cloud tasks", and the page says both once. - Keys: Week 2 "Set up Langfuse" (
anw-l13b: keys in the Git-ignored.env, never in chat; a key shown anywhere is exposed and replaced by the owner). - Langfuse:
anw-w4-l10and its details page (the standing rule, Evy'sevy.<team>.<rule>scores, the ruling filed again for a new project or key, the 30-day free-plan window), andaibl-langfusereferences/plans-retention-selfhost.md(Hobby 30 days on the checked date;export_traces.pykeeps the long-term local copy underruns/). - The daily check:
anw-w4-l08(turning it off by ruling) andanw-w4-l12Part 3 (Evy's time, the "brief me" fallback, the stop for an OK, "Run now" once, the Claude Code catch-up after sleep). - Help:
course/START-HERE.mdandcourse/reference/known-issues.md(known issues first, then one help post for Slack #agent-workforce-hr withaibl-workforce-help-post, or the student posts it). No other channel or link is named.
Not tested by us, and the page says so or sends the student to the current docs: any cloud routine, ChatGPT scheduled task or always-on server run of the student's own work; whether a cloud routine can record into the student's Langfuse; how each product signs in to connected tools.
- The Hattie B's example below is made up.
| Figure ID | Lesson | What it shows | Caption | Alt text |
|---|---|---|---|---|
w4-l13-keep-running | anw-w4-l13 Part 1 | Four boxes in a row: your computer (everything, while it's awake), cloud routines in Claude Code (jobs that only read, from your GitHub copy), ChatGPT scheduled tasks (connected tools only), an always-on computer (everything, the hardest path); under all four, one bar: "keys never in a chat" | Where your agents can run, and what each place can reach. | Four places your agents can run. Your own computer runs everything, but only while it's awake. Cloud routines in Claude Code run jobs that only read, from your copy on GitHub. ChatGPT's scheduled tasks use only the tools you connect there. An always-on computer can run everything, and it's the hardest to set up. In every case, keys never go in a chat. |
Here's what you have now. Say you own Hattie B's, a hot chicken restaurant. You have an email team that drafts replies and never sends one without your yes. The Evaluation Analyst (Evy) grades your team's work every day and tells you in your first brief how it did. A daily check keeps your organization's records straight, and Langfuse records everything your agents do. Your organization is in everyday mode, and all of it keeps running.
This page is what to do next. It covers three things: what runs while your computer is off, how to keep Langfuse wired in, and how to keep the daily check running. Then it says where to get help.
Your "what's next" page: the ways to keep your agents running, what each one can and can't do, and how to keep your checks going.
The course ends. Your work doesn't, and neither should your checks.
The Course Companion (The Professor) guides you in its own thread. Your Chief of Staff lists what runs on a schedule and saves your page. The Operator (Gigawatt) checks your recorder.
You read, ask, and decide. Nothing here changes on its own, and nothing new is set up without your yes.
Before you start
- Your organization is in everyday mode. You switched in "Move your organization from class to work". If you haven't switched yet, you can still read this page.
- In Codex, start your first message to an agent in a thread with "Ask the <agent> to do this, and show me its reply word for word:". Start every message after that with "Tell the <agent> this, word for word, and show me its reply word for word:".
Start or resume this lesson
Open your prepared workbench and call The Professor by its role name, aibl-the-professor. Paste only this.
Use the native aibl-the-professor agent to start the Week 4 page What's next, after the course. Read course/workforce/week4/prompts/anw-w4-l13.txt and course/workforce/week4/lessons/anw-w4-l13.md. Ask me one question at a time and wait. I will talk to my Chief of Staff and Gigawatt in their own threads.
The full instructions your agent follows for this page are in course/workforce/week4/prompts/anw-w4-l13.txt.
You see what can run while your computer is off, and what can't. You get a list of your own scheduled jobs and where each one could run.
Where your agents can run, and what each place can reach.
While your computer is off
Your agents run on your computer. They work only while it's awake and the app is open. When you come back, your first brief catches up on what was missed. For most people, that's enough, and you don't need anything on this page.
The ways to keep them running, simplest first
| Option | Good for | What it can't do | Where your keys live |
|---|---|---|---|
| Cloud routines in Claude Code (scheduled agents that run in the cloud) | Jobs that only read, like a nightly summary of yesterday | It runs from a fresh copy of your workbench on GitHub, at most once an hour. It can't see anything kept out of GitHub: your records in runs/, your drafts, your customer mail, your key file. It can't read your chats with your Chief. It uses your connected tools without asking, so connect only tools that read. | Not in your workbench copy: your key file never goes to GitHub. How it signs in to your connected tools: check the current docs. |
| ChatGPT's scheduled tasks (the earlier pages call them cloud tasks) | Jobs that need only the tools you connect in ChatGPT and what you upload | It can't reach your workbench folder or any file on your computer, and it can't use your agents or read your chats with your Chief. | Never in the chat. How ChatGPT signs in to your tools, and which ones a scheduled task can use: check the current docs. |
| An always-on computer (a small machine or server that stays on, like the one we run our own agents on) | Everything, all the time | Nothing is ready on it. Your agents, your tools and your recorder all have to be set up again there, with private network access. That's the hardest path. | On that server, in its own private key file, put there by you. Not in a chat, and not in GitHub. |
None of these can read your chats with your Chief, so Evy's daily check stays on your own computer. The course doesn't teach these setups. When you want to try one, try one job at a time, and start with a job that only reads.
Keep your keys out of chats
A key is like a password for a tool. On your computer, your Langfuse keys live in the file named .env in your workbench folder, the one you filled in yourself when you set up Langfuse. Git ignores that file, so it never goes to GitHub. Never paste a key into a chat, a screenshot, a note or a file you save to GitHub. If a key ever shows up in a chat, treat it as exposed: make a new one in that tool and put it in place yourself.
On the Hattie B's practice inbox
The owner wanted a summary of the day's catering emails waiting at 6 AM, before anyone opened the laptop. The owner asked the Chief which scheduled jobs could run elsewhere. The Chief listed two. Evy's daily check reads the chats with the Chief, so it stays on the laptop. The email team drafts from the workbench and waits for the owner's yes, so it stays too. A catering summary that only reads the inbox might run as a ChatGPT scheduled task, if a scheduled task can reach that inbox. The owner made a note to check the current docs first, connect the inbox to read only, and try it after the course, one job, and nothing else. This Hattie B's example is made up.
Do this now: list what runs on a schedule
In Claude Code: start a new thread, type @ and pick aibl-chief-of-staff. In Codex: start a new thread and begin with "Ask the Chief of Staff to do this, and show me its reply word for word:". Say:
List every job that runs on a schedule in my organization. For each one, tell me what it reads and whether it needs files or chats that live only on this computer. Don't change anything.
Read the list. A job that needs your chats or your workbench folder stays on your computer.
You make sure Langfuse keeps recording, and that Evy's grades keep reaching it. You get a recorder you can trust after the course.
What keeps it wired
- Keep the recorder on. It records only while it's on, and only from the moment it works. A gap in the record is a gap in what you know.
- Your standing rule. Whenever a new app, agent or computer appears, your Chief flags it, Gigawatt wires the recorder in and proves it, and the Charter Steward records it. An always-on computer is a new computer, so the same rule covers it.
- Evy's grades keep flowing back. Evy puts each grade on the record as a score named like
evy.<team>.<rule>, as long as your filed ruling "Evy writes grades to Langfuse: on" matches your Langfuse project and key. If you change projects or make a new key, Evy stops writing until the Charter Steward files that ruling again. - The 30-day window. The free Langfuse plan keeps records for 30 days. After that, your earliest records drop off.
- The local copy is your long-term history. Evy keeps a copy of what it needs for grading on your computer, in the job's folder under
runs/. That copy is what Evy quotes, and it stays after Langfuse's window has passed. It holds what grading needed, not every record. Plans and limits can change, so check the current Langfuse docs before you rely on a number.
On the Hattie B's practice inbox
Three weeks after the course, the owner made a new Langfuse key after a laptop repair. The next brief said Evy's grades had stopped reaching Langfuse. The owner asked the Chief to have the Charter Steward file the ruling again for the new key, and the next daily check read its scores back.
Do this now: ask Gigawatt to check the recorder
In Claude Code: start a new thread, type @ and pick aibl-gigawatt. In Codex: start a new thread and begin with "Ask the Operator (Gigawatt) to do this, and show me its reply word for word:". Say:
Check that Langfuse still records everything my agents do in every app I use, and prove it by reading a test record back. Tell me whether my ruling for Evy's grades still matches my Langfuse project and key, and where my long-term copy of the records lives. Don't show me any key.
What you should see
For each app, whether it passed, with a link to the test record. Whether Evy's ruling still matches. Where the local copy lives. No key anywhere in the reply.
You make sure Evy's daily check keeps running, and you learn how to stop it. You get a check you know how to look after.
What keeps it running
- Evy's time. If you picked a time in "Move your organization from class to work", your morning digest starts Evy's check at that time, while your computer is awake and the app is open.
- "Brief me" is the fallback. Your first brief of the day runs Evy's check if it hasn't run yet. It never runs twice in a day. If your computer was asleep at your time, Claude Code runs one catch-up when it wakes.
- When a scheduled run stops to ask for your OK. It doesn't guess. It stops, saves a short note saying what it was waiting for, and your next brief tells you. Open the morning digest in your app, click "Run now" once while you watch, and approve what it asks. Then the next morning run usually has what it needs. A tool that asks for your OK every time will stop the run every time; ask your Chief to help you take that step out of the morning run.
- Turning things off. Say "Turn off Evy's daily check" to your Chief and have the Charter Steward record your ruling. The latest ruling wins, and your earlier rulings stay on file. To stop the morning digest itself, ask your Chief to walk you through pausing it in your app, one click at a time.
On the Hattie B's practice inbox
The owner went on a week's holiday with the laptop shut. Nothing ran that week, and nothing broke. On the first morning back, the owner said "brief me". The brief said what was missed and what it caught up on, and Evy's check ran once.
Do this now: check when it last ran
In your Chief of Staff thread, say:
When did Evy's daily check last run, and what started it: my set time or my first brief? If a scheduled run stopped to ask for my OK, show me what it was waiting for.
What you should see
The day and time it last ran, and whether your set time or your first brief started it. If a run stopped for your OK, the step it waited on.
If you get stuck
| What you notice | What to do |
|---|---|
| You want work to run while your computer is off | Read Part 1 again, and pick one job that only reads. Ask The Professor, in its thread, what each option can and can't do. |
| Something asks you to paste a key into a chat | Stop. Keys go in a private key file, put there by you. If you already pasted one, make a new key in that tool and put it in place yourself. |
| Gigawatt's check says "sent" but not "read back" | It hasn't passed. Ask Gigawatt to read it back. |
| Evy's grades stopped reaching Langfuse | Your project or key changed. Ask your Chief to have the Charter Steward file the ruling "Evy writes grades to Langfuse: on" again. |
| Evy's check didn't run at your time | Say "brief me": your first brief runs it. Your computer has to be awake and the app open at your time. |
| You want everything to stop | Say "Turn off Evy's daily check" to your Chief, and have the Charter Steward record your ruling. |
The three options, your keys, Langfuse and the daily check in more detail: more detail on this page.
Where to get help
- Check the known issues first. Ask The Professor, or open
course/workforce/reference/known-issues.mdin your workbench. Each fix there has been checked. - Then ask in Slack. If nothing there fixes it, The Professor can help you write one help post for the Slack channel #agent-workforce-hr, or you can post there yourself.
- Never put a key, a password or a customer's details in a help post.
Do this now, before you go: save your page
In your Chief of Staff thread, say:
Save a copy of my "what's next" page, from course/workforce/week4/lessons/anw-w4-l13.md, where I'll find it, and tell me where it is.
What you'll keep
Your "what's next" page, saved where you'll find it; a list of what runs on a schedule and where each job could run; a recorder you've checked; and a daily check you know how to look after and how to stop.
Before you finish
Self-assessed. No gate.
Words you now own
- Cloud routine: a scheduled agent in Claude Code that runs in the cloud, from your copy on GitHub.
- Scheduled task: a job ChatGPT runs at a set time, with only the tools you connect there.
- Always-on computer: a small machine or server that stays on, so your agents can run all the time.
- Long-term history: Evy's copy, on your computer, of what it needed for grading, kept after Langfuse's 30 days.
Next: You've finished the course. Keep your first brief each morning, and keep grading as much as your check-by-hand step says.