Write a goal an agent can finish, and see the loops that run your teams
What tells an agent it’s done, and what starts it again?
- Source
- Internal's lesson, src/sara-week3/internal/lessons/anw-w3-l13.md (Internal b518c35b, 2026-10-22 17:15), read by md_week3.py.
- Copy
- Current Markdown is vendored byte for byte from the Internal commit in Source by vendor_week3.py. Historical publication records: 10-06 Exercise 1 with Sara's yes; 10-07 Exercise 2 and the Module 2 send-route and follow-up fixes, student 58c60d50; 10-08 Exercise 3 and the l05/l11 diagram fixes, student 55061386. Those records do not qualify later Outlook changes. Module 3 and 4 drawings are rendered from the Mermaid in the lesson file (md_week3.py --figures).
- Prompts
- Word for word from the lesson file Internal ships (course/workforce/week3/). 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 scroll (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 Week 3 Module 1, Write a goal an agent can finish, and see the loops that run your teams. Read course/workforce/week3/lessons/anw-w3-l13.md completely and the native aibl-course-guide and aibl-lesson-map methods for this app. This lesson teaches, and saves one thing: the owner's finished goal, as this lesson's dated entry in work/course/staff/progress.md, shown first and written only on their yes, with every earlier entry unchanged. Do not edit, stage or propose an edit to any other file. Do not start, schedule or cancel a loop, a goal or a job.
Use these words exactly. The loop (Week 1): look, think, act, check, then finish or stop. A loop that runs again has these parts: what starts it, the goal, the work (look, think, act, check), the receipt, then it stops or waits for the next start. The five parts of a goal: task, why, outcome, constraints, verification; then a sixth part, bound (stop after N turns, minutes or dollars), and a seventh line, if stuck (stop and tell me when ...). The four kinds of loop: turn-based, goal-based, time-based, proactive.
Part 2. When the owner asks, find one work order from a job the owner ran in Week 1 or Week 2 (runs/<job>/work-order.md, or work-orders/), read only, and show its goal, done and never or boundaries lines exactly, with the file path. If there is none, say so and where you looked; never invent one. Then open .claude/agents/aibl-chief-of-staff.md, read only, and quote exactly the sentence that lists the careful work order's seven parts and the sentence "If you cannot fill the form the job needs, do not send the job." Then open .claude/agents/aibl-gigawatt.md, read only, and quote step 4 of its "How you work" section ("Run the check") exactly. Explain in plain words how the job and agent, goal, done and boundaries line up with task, outcome, verification and constraints. The work order says why each agent was picked; it has no line for what the result is for, and none for a bound.
Part 4 is practice. Ask open questions only in this part: no numbered choices and no example answers, whatever your usual style. The owner has written a goal for one job of their own and judged it against the seven tests on paper. Ask them to read it to you, then ask which tests they judged it to pass and fail. Then, for each of the seven lines in order, say whether you read it as meeting that line's test and quote the words you are judging. Where you disagree with the owner's own judgment, say so and why, once. Never write, rewrite or complete any line of the goal, and never offer a better version. If the owner asks you to write it, say that writing it is the practice and ask which line they want to try again. When the owner says the goal is final, save it word for word, with their marks, as this lesson's entry in work/course/staff/progress.md: show the text, get their yes, write, read it back. The Architect (Archie) reads it first in the lesson "Design a team for your own work, with Archie".
Staff only: review scope, dependencies, changes. Not student-facing.
Status (8 October 2026, tier 1 course lane). Student text and controlling prompt updated on the 8 October research synthesis (tyler-w3-ex3/research/SYNTHESIS.md; COURSE-FIT rows 5 and 11). The goal gains a sixth part, bound, and a seventh line, if stuck; the five parts, their examples and the judged example are kept. The lesson now saves one thing: the finished goal, as this lesson’s dated entry in work/course/staff/progress.md, shown first and written on the student’s yes (the shape the Professor’s pen already reaches for progress notes). The paste to Archie in “Design a team for your own work, with Archie” opens with it. Earlier state: Sara’s sign-off draft of 1 October; the /loop demo cut by Sara (kept in commit 6efaed6c); the agent claims were checked against internal origin/main on 1 October and can go stale, so re-check before shipping.
Sources. Sara’s brief (1 October). Camp: camp-hq/curriculum/goal-driven-authoring/week-3-goal-shaped-tasks-teaching-moment.md (task, why, outcome, constraints, verification) and agent-native-camp/weeks/week-04/w4-01-one-front-door.md (trigger, goal, work, receipt, sleep; the four kinds), rewritten, not copied. Week 1 anw-l01.md (the loop; a loop does not mean unattended forever). Agent files: aibl-work-order/SKILL.md line 8 (not shown to students; the Professor cannot read .claude/skills/), aibl-chief-of-staff.md line 64, aibl-gigawatt.md step 4. Claude Code scheduled-tasks docs and Codex /goal (1 October). The bound and the stuck line: Claude Code /goal (“or stop after 20 turns”; the Impossible verdict), Devin’s three-hour split, Copilot’s acceptance criteria.
Open items. The Professor’s file (line 78) still says the organization has no schedule; students who did anw-w2-b01 have one. Camp’s w4-01 tells students Escape stops a /loop 10m; per the docs it does not. Part 3’s goal-based claim covers the background build team (spine v2). Lesson and prompt files must reach course/workforce/week3/; dep-13 applies.
Every agent works through a job in a loop: look, think, act, check. You met it in Week 1.
Today you learn what makes a loop stop: a goal with a check and a bound. You write one for a job of your own, judge it, and keep it for Archie.
You also see which kind of loop each of your three teams uses this week.
How a loop works and what makes it stop. How to write a goal an agent can finish. A goal for one of your own jobs, saved for the Architect (Archie).
Each team this week runs with less of you watching. A clear goal tells your agents what done looks like, and when to stop even if it isn’t. A done check is a test anyone could repeat to prove it’s finished, even when you weren’t watching.
The Course Companion (The Professor) shows you a work order and the agents’ files, gives a second opinion on your goal, and saves it on your yes.
You write the goal and judge it. Nothing is run.
Before you start
Have paper and a pen, and the work stream you named earlier this week in mind.
Start or resume this lesson
Open your prepared workbench and call the Course Companion (The Professor) by its role name, aibl-the-professor. Paste only this.
Use the native aibl-the-professor agent to start the Week 3 lesson "Write a goal an agent can finish, and see the loops that run your teams". Read course/workforce/week3/prompts/anw-w3-l13.txt and course/workforce/week3/lessons/anw-w3-l13.md. Ask me one question at a time and wait. Change nothing in my workbench.
The loop you met in Week 1
In Week 1 the Communications Writer drafted a workshop invitation for you. This is how it worked:
- Look: it read the two notes.
- Think: it saw they disagreed about the date, and chose to leave a gap.
- Act: it wrote the draft.
- Check: it checked each fact against the notes.
Then it finished. That’s the loop.
It runs inside one job, as many times as the job needs.
It doesn’t go round forever. It finishes when the job is done, or it stops and comes back to you: at its bound, or when it needs information, permission or help.
A loop that runs again
Some loops start again after they finish.
How a loop that runs again works.In the middle, highlighted, is the Week 1 loop: look, think, act, check, as many times as the job needs.
Diagram: How a loop that runs again works. In the middle, highlighted, is the Week 1 loop: look, think, act, check, as many times as the job needs.
The goal is what makes a loop stop. If no one has said what done means, the agent can’t tell when it has got there.
The receipt is how you check a loop you didn’t watch. Without one, you’re only hoping it worked.
You saw receipts in Week 1: the work order, the draft and the review, each saved in the job’s folder.
The same shape runs all three of your teams this week. What changes is what starts the loop, and how much of it you watch.
Next part: The parts of a goal an agent can finish
Five parts, a bound and a stuck line
A goal an agent can finish says five things, then sets a bound and a way out.
| Part | What it says | Example |
|---|---|---|
| Task | What work the agent owns, in one job. | Draft this week’s post about a client win. |
| Why | What the result is for. It helps the agent make good calls when something is unclear. | So people who might hire me see work like theirs. |
| Outcome | What is true when it’s done. The result, not the steps. | A post under 200 words that I could publish today. |
| Constraints | What must not happen, and what waits for your yes. | Use only the facts in the project notes. Name the client only if they’ve said yes in writing. Nothing is posted. |
| Verification | A done check: a test that proves it’s finished. Someone else could repeat it and get the same answer. | Every fact in the post points to a line in the project notes. Anything missing shows as a visible gap. |
| Bound | When to stop even if it isn’t done: so many turns, minutes or dollars. | Stop after 20 turns or 10 minutes. |
| If stuck | One line: stop and tell you, and when. | If stuck: stop and tell me when the notes don’t cover a fact the post needs. |
Take only the task: “Draft this week’s post about a client win.” The agent would still have to guess what the post is for, how long it should be, whether it can name the client, when it’s finished, and when to give up.
A goal is specific about where the work ends. The agent decides the steps.
You’ve seen most of this before
Your Chief of Staff writes a work order for every job. You read one before you said go in Week 1. Four of the five parts of a goal are already in it.
A work order like the one from Week 1, shortened.Each tag shows which part of a goal the line is. There is no line for why.
for: Communications Writer
A workshop invitation exists, in your voice, that a reader could act on the moment the date and link are filled in, and the Quality Reviewer has passed it.
Both gaps still show, [DATE TO CONFIRM] and [REGISTRATION LINK NEEDED], and every fact points to a source note.
Never state a date, a price or a link. Never send or post. One revision if the Quality Reviewer finds a problem, then hold for you.
The notes it may use, what comes back, your one decision, where it’s filed.
Diagram: A work order like the one from Week 1, shortened. Each tag shows which part of a goal the line is. There is no line for why, and none for a bound.
The Chief of Staff says why it picked that agent. It doesn’t say what the result is for, or when to stop. The why and the bound are yours to give.
The Chief of Staff’s file says it plainly: if it can’t fill in the work order, it doesn’t send the job.
A loop that knows when to stop
The Operator (Gigawatt) builds agents. When it finishes one, it runs a check that prints a pass or a fail. If the check fails, it fixes one thing and runs the check again. After one repair, it stops and brings the next failure to you.
That’s a goal with two ways to end. The check passes, or the stop rule hands it back to you. That rule is a bound and a stuck line in one: one repair, then tell you.
Do this now
Find the five parts in one of your own work orders
The Professor opens a work order from one of your Week 1 or Week 2 jobs, then the Chief of Staff’s and Gigawatt’s files. You say which part of a goal each line matches before it tells you.
Thread: your lesson thread with The Professor. Say this:
Show me one work order from a job I ran in Week 1 or Week 2: its goal, done and boundaries lines, word for word, and where the file is. Then show me the line in the Chief of Staff's file that says when it won't send a job, and Gigawatt's step that runs the check. Let me say which part of a goal each line matches before you tell me.
Next part: The four kinds of loop, and where each of your three teams uses them
The four kinds of loop
Loops differ by what starts them again. There are four kinds.
| Kind | What starts it again | Where you see it this week |
|---|---|---|
| Turn-based | You say something. It answers, then waits for you. | Your conversations with The Professor, and with Archie, which asks one question at a time and waits. |
| Goal-based | It keeps going until the done check passes, or its stop rule hands it back to you. | Every job your Chief of Staff routes, on all three teams. The Operator (Gigawatt) checking each new agent it builds. |
| Time-based | The clock. | Your morning email collection, if you did the Week 2 bonus, and your email department’s inbox check once you plan it and its live test passes. You set every schedule yourself. Your organization never schedules itself. |
| Proactive | Something happens, like an email arriving. | Not in this course. Every job this week starts with you. |
Across your three teams, you watch less and less as it happens:
- Customize: you run three observational evals on your own email and read every draft before anything goes out.
- Build: you test it yourself on five cases, and read what each agent did.
- Delegate: you’re away while it’s built. You see only the receipts when you come back.
The less you watch, the more you rely on a clear goal and its receipts.
A goal is not a schedule. A goal says when the work is done. A schedule says when the work starts. A loop can have one, the other, or both.
Next part: Write a goal for one job of yours, and judge it yourself
Write a goal for one job of yours
When Archie designs a team, it decides what a good result looks like before it picks any agents: “The checks you set here decide the seats, not the other way round.” Seats are its word for agents.
Pick one job from the work stream you named earlier this week. One run of it, not the whole work stream.
Write the goal on paper, one line each. These questions help.
- Task: What’s the one piece of work? What starts it?
- Why: Who receives the result? What would go wrong if it were late or wrong?
- Outcome: Picture it finished and in front of you. What is it, and how long or how big?
- Constraints: What must it never do? Which of your decisions from earlier this week does it touch?
- Verification: What could you count, compare or look up to know it’s done?
- Bound: How many turns, minutes or dollars is one run worth?
- If stuck: What would mean it can’t finish? Write: “If stuck: stop and tell me when ...”
Judge it against the seven tests
Mark each line pass or fail, and say why in a few words.
| Part | The test |
|---|---|
| Task | Is it one job? Could you tell when it starts and when it’s over? |
| Why | Does it say who the result is for, or what it changes? |
| Outcome | Does it describe the finished result, not the steps to get there? |
| Constraints | Does it name what must not happen? Does it say what waits for your yes? |
| Verification | Could someone else run the check and get the same answer? “Looks good” fails. |
| Bound | Does it say when to stop even if it isn’t done? |
| If stuck | Does it say to stop and tell you, and when? “Keep trying” fails. |
Here is a goal judged this way, then rewritten, for replying to a client who wants to move a session.
| Part | First try | Mark | Rewritten |
|---|---|---|---|
| Task | Reply to a client’s request to move Thursday’s session. | Pass One job, with a clear start and end. | No change. |
| Why | (none written) | Fail The agent can’t weigh a trade-off. | Keep the booking. A client who can’t find a new time may cancel. |
| Outcome | Handle it. | Fail That’s a step, not a result. | A reply offering two open times this week, ready for me to send. |
| Constraints | Don’t offer a time I’m teaching. Nothing is sent. | Pass Names what must not happen and what waits for you. | No change. |
| Verification | Make sure it’s good. | Fail An opinion, not a check. | Both times are open on my calendar. |
| Bound | (none written) | Fail It could search all day. | Stop after 15 minutes. |
| If stuck | (none written) | Fail No way out. | If stuck: stop and tell me when fewer than two times are open. |
Verification fails most often. “It reads well” is an opinion. “Every date matches the calendar” is a check. A bound and a stuck line are usually just missing.
Do this now
Write it and judge it
Write your goal on paper, then mark each line against its test.
Get a second opinion, then save it for Archie
Read your goal and your marks to The Professor. It says which tests it reads each line as meeting, and where it disagrees with you. It doesn’t write any of it for you. When you say it’s final, it shows you the text and, on your yes, saves it word for word as this lesson’s entry in work/course/staff/progress.md. Archie reads it first when you design your team, in the lesson “Design a team for your own work, with Archie”.
Thread: your lesson thread with The Professor. Say this:
I've written a goal for one job and marked it against the seven tests. I'll read you both. For each line, tell me whether you read it as meeting its test, and where you disagree with my mark. Don't write or rewrite any of it. When I say it's final, show me the text and save it word for word as this lesson's entry in my course progress note, on my yes.
Next part: You're done
Where to look when it goes wrong
| What you notice | First question |
|---|---|
| The Professor rewrote my goal, or offered a better one | Ask it to stop. Writing it is the practice. Ask it only which tests each part meets. |
| I can’t think of a check for verification | Ask yourself how you’d know, without reading it, that a draft was finished. A count, a list that matches, every fact pointing to a source. |
What you’ll keep
Your goal for one job, with your marks, saved as this lesson’s entry in work/course/staff/progress.md. Archie reads it first when you design your team.
Before you finish
You can say how a loop that runs again works, which kind runs each of your three teams, and which two parts of a goal your Chief of Staff’s work order leaves to you.
You’ve written a goal for one job, judged it yourself, and saved it for Archie.
Next: Set up your email department. You set up the team built for you, and see the roles and handoffs in a working agent team.
Words you now own
- what starts a loop
- task, why, outcome, constraints, verification, bound
- if stuck
- done check
- stop rule
- turn-based
- goal-based
- time-based
- proactive
On your screen when this ends · 3 checks
- I've written a goal for one job, marked each of its seven lines pass or fail, and saved it for Archie
- I can say which kind of loop runs the jobs on my three teams, and how much of each I watch
- I can say what makes a loop stop, and what a receipt lets me check
Self-assessed. No gate.