Make sure everything your agents do is recorded
Is everything my agents do recorded, in both apps, and will it stay that way when I add something new?
- Source
- Internal's lesson, src/week4/internal/lessons/anw-w4-l10.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 lesson Make sure everything your agents do is recorded. Read course/workforce/week4/lessons/anw-w4-l10.md completely, its details page course/workforce/week4/reference/anw-w4-l10-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 install, change or test a recorder, read or handle a key, or record a ruling for the student. The student talks to the Operator (Gigawatt) and to the Chief of Staff 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. Langfuse is the recorder the student set up in Week 2: it keeps a record of every turn an agent takes, every tool it uses, and every agent it starts, linked to the agent that started it. A complete record says which job, which team, which agent, which app (Claude Code or Codex) and which computer, and which version of the recorder and of the agent files, with no email address and no raw computer name or folder path. Gigawatt checks each app the student uses: it sends one test record with a planted fake secret, reads it back by its id, confirms every field and the link to its parent, confirms the fake secret is hidden, gives the student the link to the record, then runs the catch-up sweep and confirms nothing was sent twice. Gigawatt never overwrites or removes a recorder the student already has: it reports any existing one, installs beside it under its own names, and removes only its own entries. The free Langfuse plan keeps records for 30 days; Evy keeps a local copy for grading. The standing rule: whenever a new app, agent or computer appears, the Chief of Staff flags it, Gigawatt wires in the recorder, sets its tags and runs the read-back check, and the Charter Steward records it. Name agents job title first, nickname in brackets on first use. 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.
Evy finds the job in Langfuse but quotes only from the local job copy. After the owner asks the Chief to have the Charter Steward file "Evy writes grades to Langfuse: on", Evy writes one PASS or FAIL score per rule on the turn that did the work. The Steward files the Langfuse project and key fingerprint. If either changes, the ruling must be filed again. HOLD never goes to Langfuse. Evy reads every score back and checks a second send made no duplicate (except near midnight UTC, when the second send is skipped and the single score read back is the proof). In a blind grading sample, Evy's scores wait until the owner seals a grade, then both sets go in together. Score Analytics shows the comparison; the saved agreement count remains the count of record.
Ask open questions only.
Part 1. If asked, explain what a complete record says, from the page.
Part 2. If asked, read Gigawatt's check result the student shows, read only, and say back plainly which app passed, which field is missing, and whether an existing recorder was found and left alone. A check is passed only when the record was read back.
Part 3. If asked, help the student find the job, team, agent, app and computer on the record they opened. Also read Evy's score read-back line and the score the student opened. Explain a blind hold without showing Evy's grade before the owner's is sealed.
Part 4. If asked, read the standing rule the student shows, read only, and say back what happens the next time a new app, agent or computer appears.
Otherwise, if asked, explain the step and what the student should see.
Staff only: status, sources, what the build supports. Not student-facing.
Status: revised 12 October 2026 from the recorder lesson draft and the Evy Langfuse score-write lane. The original recorder spec came from spine v2 Module 5 and DESIGN-week4-2026-10-08.md section 10 (student spec: trace name, session, userId never an email, environment course, release and version; metadata harness, machine, workbench, team, job, seat, parent links; short tags, no raw hostnames or paths; the Chief writes .aibl-local/current-job.json; seeded ids, 15-minute sweep; living with other recorders; the five-step check with a planted fake secret; the new-hire rule; the 30-day free-plan window and the export).
Built in workforce/runtime/langfuse_evy.py. Evy finds a job's live records in Langfuse, quotes only from the local job copy, and writes each PASS or FAIL rule back as a score after the owner's ruling is filed. The script checks the project and key fingerprint, reads scores back, and checks a repeat does not double them. Earlier lessons this week do not depend on this one: Evy grades from the job folders and writes "not verified" where only a recording would show something (PR #202, Evy's read item 4), so a student without Langfuse reaches this lesson with grades in hand. The key file is the Git-ignored .env at the workbench root, the one Week 2's "Set up Langfuse" uses.
- The Langfuse lane owns the recorder, the check and the Codex-on-Mac proof. The page has the student ask Gigawatt in plain words and names no command.
- Week 2's
anw-l13bset up Langfuse in Claude Code and in Codex (each app the owner uses, set up once). A student who noted "no recording set up" in Week 2 starts here: the page's stuck table sends them to Gigawatt's setup, which this lesson's check then proves. - The standing rule: the page has the student ask the Chief to file it as a ruling through the Charter Steward. The Langfuse lane decides where it lives in the Chief and Gigawatt (Tyler: the new-hire rule goes into the student Gigawatt and Chief).
- The Hattie B's record below is a made-up worked example.
| Figure ID | Lesson | What it shows | Caption | Alt text |
|---|---|---|---|---|
w4-l10-record | anw-w4-l10 Part 1 | One record drawn as a card with five labelled rows (job, team, agent, app, computer) and an indented child card "an agent it started" linked to it; a stamp "fake secret: hidden" | Every record says which job, which agent, which app and which computer, and hides anything secret. | One record in Langfuse says which job it belongs to, which team, which agent, which app and which computer. The record of an agent started by another agent is linked to the one that started it. Anything secret in it is hidden. |
In Week 2 you set up Langfuse, the recorder that keeps a record of everything your agents do. This week Evy finds the job's records there and grades from a local copy on your computer. Every record should say which job, which agent, which app and which computer it came from. With your filed ruling, Evy also puts each grade on the record for the turn that did the work.
Today the Operator (Gigawatt) checks the recorder in every app you use, and fixes what's missing. It never replaces a recorder you already have. Then you add one standing rule: whenever a new app, agent or computer appears, it gets wired in.
Langfuse recording your whole organization in both apps, and a rule that keeps it that way.
Evy can only grade what was recorded. A gap in the record is a gap in what you know.
The Course Companion (The Professor) guides you in its own thread. Gigawatt checks and fixes the recorder. Your Chief of Staff watches for anything new, and the Charter Steward records your rule.
You read the check result, open one record and its grades yourself, and ask your Chief to file the two rulings through the Charter Steward.
Before you start
- Langfuse is set up. You did it in Week 2, in "Set up Langfuse". If you noted "no recording set up" then, Gigawatt sets it up now; you'll need your Langfuse sign-in.
- The apps you use. Claude Code, Codex, or both. Gigawatt checks each one.
- 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 lesson Make sure everything your agents do is recorded. Read course/workforce/week4/prompts/anw-w4-l10.txt and course/workforce/week4/lessons/anw-w4-l10.md. Ask me one question at a time and wait. I will talk to Gigawatt and my Chief of Staff in their own threads.
The full instructions your agent follows for this lesson are in course/workforce/week4/prompts/anw-w4-l10.txt.
You see what every record should carry. You get what to look for in Part 3.
Every record says which job, which agent, which app and which computer, and hides anything secret.
On the Hattie B's practice inbox
When Evy graded the kids plate question, Evy found the job's records in Langfuse. The record said "agent: the writer", so Evy could connect the line about the fries to the writer, not just the email team. Evy quoted the local copy of the job to support the grade, then put the writer's rule score on the writer's last record for that job. This Hattie B's example is made up.
Records don't keep forever
The free Langfuse plan keeps records for 30 days. Over a few weeks, your earliest records drop off. Evy keeps a copy of what Evy needs for grading, on your computer, so a grade never depends on that window.
Living with a recorder you already have
If you already record your agents some other way, Gigawatt tells you what it found and leaves it alone. It adds its own recorder beside yours and never replaces or removes anything it didn't put there.
Do this now: find one record
If you set up Langfuse in Week 2, open it in your browser, open your project, and click any one record from this week. Which of the five lines from the picture can you find on it? Keep the answer for Part 2. If you didn't set Langfuse up, skip this; Gigawatt sets it up with you in Part 2.
You ask Gigawatt to check the recorder in each app. You get a plain result for each one, proven by reading a test record back.
How the check works
For each app, Gigawatt sends one test record with a made-up secret planted in it. Then it reads that record back from Langfuse and checks every line is there, each started agent is linked to the one that started it, and the made-up secret is hidden. It also checks nothing was sent twice. A check passes only when the record was read back. A record that was only sent doesn't count.
Do this now: ask Gigawatt to check
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 this, naming the apps you use:
Check that Langfuse records everything my agents do in <Claude Code, Codex, or both>, with which job, team, agent, app and computer on every record. Tell me first if you find any recorder I already have, and leave it alone. Fix what's missing, then prove each app by reading a test record back, and give me the link.
What you should see
For each app: whether it passed, which lines were missing and are now fixed, whether you already had a recorder (and that it was left alone), and a link to the test record. No key or secret appears anywhere in the reply.
You look at one record in Langfuse. You get to see for yourself that it says what it should.
Do this now: open the link
Open the link Gigawatt gave you, in your browser. Find the five lines from the picture: job, team, agent, app, computer. Then open the recording of any job your team ran today, and check the same five.
If one is missing, tell Gigawatt in the same thread which one.
Let Evy put grades on the record
Ask your Chief of Staff to have the Charter Steward file your ruling, "Evy writes grades to Langfuse: on". The Steward records the Langfuse project and a fingerprint of the key, never the key itself. This is the one approval for Evy's grade writes. If you change projects or keys later, Evy stops writing until the Steward files the ruling again.
Do this now: check one graded job
In your Chief's thread, say:
Have the Charter Steward file my ruling, "Evy writes grades to Langfuse: on", for this Langfuse project. Then ask Evy to put the grades for one job I have already had graded into Langfuse. Show me Evy's read-back result and the link to a graded record. Do not show me Evy's grades for a job I still plan to grade blind.
Open the graded record in your own Langfuse project. Find an evy.<team>.<rule> score, such as evy.customer-support.A2, and its reason and short quote from the local job file. True means PASS; False means FAIL. HOLD never goes to Langfuse. An agent's rule belongs on that agent's last record in the job; a job rule belongs on the first Chief of Staff record. There is one score per rule, with no total or average. Evy should report a line such as "Langfuse: 4 of 4 scores read back" and confirm a repeat made no duplicate. "Sent" alone is not proof.
In Score Analytics, you can later compare evy.<team>.<rule> with your own owner.<team>.<rule> scores. The saved agreement count remains the count of record. If this job is in your blind grading sample, Evy's scores wait until you save and seal your own grade. Then both sets of scores go in together. For a job you plan to grade yourself, leave Evy's folder closed until your grade is sealed.
You make one rule for everything new from now on. You get an organization that keeps its record complete without you remembering to.
On the Hattie B's practice inbox
The student added Codex on a second laptop the week after the course. The next morning's brief said: "A new app and computer appeared: Codex on the laptop. It isn't recorded yet." Gigawatt wired the recorder in, proved it with a test record, and the Charter Steward recorded the new computer. The student read one line in a brief and did nothing else.
The standing rule
Whenever a new app, a new agent or a new computer appears in your organization:
- Your Chief of Staff flags it in your brief.
- Gigawatt wires the recorder into it and sets its job, agent, app and computer lines.
- Gigawatt runs the same read-back check.
- The Charter Steward records it.
Do this now: say yes to the rule
In a new thread with your Chief of Staff, say, in your own words:
From now on, whenever a new app, agent or computer appears, flag it, have Gigawatt wire it into Langfuse and prove it, and have the Charter Steward record it. Have the Steward record this as my ruling and show it to me word for word.
What you should see
Your ruling, word for word, filed by the Charter Steward with today's date.
If you get stuck
| What you notice | What to do |
|---|---|
| You didn't set up Langfuse in Week 2 | Tell Gigawatt in its thread. It sets it up with you: you sign in to Langfuse and paste the keys it shows into the file named .env in your workbench folder yourself. Never paste a key into a chat. |
| Gigawatt says it found another recorder | Good: it found it and left it alone. Read what it found, and ask if you're unsure what it is. |
| A check says "sent" but not "read back" | It hasn't passed. Ask Gigawatt to read it back. |
| You see a key or password anywhere in a reply | Stop and tell Gigawatt. Treat that key as exposed: make a new one in Langfuse and put it in yourself. |
| One app passes and the other doesn't | Keep using the app that passes, and ask Gigawatt what stopped the other. Nothing else this week depends on both. |
| Evy says the ruling must be filed again | The Langfuse project or keys changed. Ask the Chief to have the Charter Steward file the ruling again for this project. Never paste a key into chat. |
| Evy says "held until you grade it" | Save and seal your own grade first. Evy's scores and yours can then go to Langfuse together. |
How the check works and what each record carries: more detail on this lesson.
What you'll keep
Langfuse recording your whole organization in every app you use, proven by a test record read back, with any recorder you already had left alone; a filed grade-write ruling and one graded record you checked; and a standing rule that wires in anything new.
Before you finish
Self-assessed. No gate.
Words you now own
- Record: what Langfuse keeps for each turn an agent takes, with its job, team, agent, app and computer.
- Read back: fetching a test record from Langfuse to prove it arrived complete.
- Standing rule: your ruling that anything new gets wired into the recorder.
Next: Build your proof page.