Agent Lab / Test Class
Try a lesson on four readers before it meets real ones
Most training gets tested on the people it was written for, after it ships. This is a rehearsal. You give it a topic. One AI writes a short lesson with a quiz. Four more AIs read it, each playing a kind of reader I meet at work: the executive who skims, the new hire who doesn’t know the jargon yet, the colleague reading in her second language, and the expert who checks every claim. Each hands back a short report card. A fifth AI, the lead, reads the four cards, decides what to fix, sends the lesson back for a rewrite and runs the class again. Three rounds at most. Then it stops and waits for a person to say yes or no.
Everything on this page happened. The three runs below are replayed from the files the agents saved, word for word.
The four readers
Four readers who can’t compare notes
If the four students could see each other’s answers, they would drift toward one opinion, and I’d have one reader instead of four. So each student gets the lesson and nothing else: no answer key, no other student’s card, no earlier round. The writer never hears who said what; it gets a list of fixes about lines and questions. The lead sees everything, and it is the only one that does.
Dana is the odd one. Tell an AI to skim and it still reads every word; it just acts the part. So the lead hands her a copy with only the title, the headings, the bold lines and the first sentence of each paragraph. That is what a real skimmer sees, and she can only answer from that.
| Who | The topic | The lesson | Dana’s skim copy | The answer key | The report cards | The fix list |
|---|---|---|---|---|---|---|
| The leaddecides, never writes | Sees | Sees | Makes it | Grades with it | Reads them | Writes it |
| The writerdrafts and rewrites | Sees | Writes it | Writes it | Sees | ||
| Danaan executive who skims | Sees only this | |||||
| Samfive days into a new job | Sees only this | |||||
| Lucíareading in her second language | Sees only this | |||||
| Theothe expert who checks every claim | Sees only this |
How it runs
The lead runs the class. The code holds the rules.
At work, my course agents ran one after another, like a relay: when one finished, the next started. The lead here does something different. It reads what came back and decides what happens next, which is what people mean when they say one agent is orchestrating others. The rules it can’t bend are written as code rather than instructions: how the skim copy is made, how the quiz is graded, what counts as a pass, three fixes at most, and no names to the writer.
- 1. The lead briefs the writerThe topic. After round one, the fix list too.
- 2. The writer draftsA short lesson, a five-question quiz and a separate answer key.
- 3. Code checks the draftFive questions, under the word limit. It also cuts Dana’s skim copy.
- 4. Four students read it at onceThree get the lesson. Dana gets her skim copy. None gets the answer key.
- 5. Code marks it, the lead decidesThe code grades the quiz and applies the pass rule. The lead writes up to three fixes.
- Did not passBack to step 1 with the fix list. Three rounds at most, then it stops anyway.
- Passed, or out of roundsThe lead writes a report and stops. A person approves it or says not yet.
- On my laptop
- Six text files, each a job description in plain English: one for the lead, one for the writer, one for each student. Claude Code reads them and runs the class, starting the four students in one go so they work at the same time. Every lesson, report card and decision is saved as a file, so the folder from a run is its own record.
- On this site
- The same class rebuilt as a program, so the rules can’t be argued with and you can run it live. Each round is five calls to the AI: one writer, four students. The lead’s decision is a sixth. Everything the AI returns has to fit a fixed form, or the code sends it back.
- Where it stops
- Three rounds on my laptop, two here, a few minutes in all. Then it writes a report and waits. It has no way to send, post or publish anything.
Try it
- Amber An AI wrote this: the lesson, the report cards, the lead’s decision.
- Blue A person decides here.
- Underlined This sentence changed since the last round.
Pick a run. You’ll see the lesson on the left, the four report cards on the right, and what the lead decided. Press play to watch it round by round.
Under the hood
What the AI writes and the software checks
Written by Claude
- The lesson and its quiz, and every rewrite from the fix list
- Each student’s report card, in character
- The lead’s decision: what to fix and why
Checked by the software
- Dana’s skim copy is cut by code, not by asking the AI to skim
- Students never see the answer key; the code marks the quiz
- The fix list holds at most three items and never names a student
- If everyone scores 5 of 5 in round one, the quiz was too easy: that round fails and the writer must make it harder
- The class stops after three rounds (two in live mode), whatever the scores
- A topic that names a person is refused before anything runs
- Nothing is sent or published; the run ends waiting for a person
These checks are tested automatically whenever the Lab is updated.
See the code
What it is made of
The laptop version has no program in it at all. It is six text files and a folder of results, and anyone with Claude Code can download it and run it. The student files are the interesting part to change: each one describes a reader in a page of plain English, so you could replace my four with the readers you actually have.
On GitHub:The six files and the three runsWhat I learned building it
The students are AI standing in for four kinds of reader. They catch what those readers tend to catch, not what a real class would say. Use it before a pilot, not instead of one.