Lab: Instrument One Manual Workflow, Your Minutes Baseline
Everything in this semester about whether an agent "helped" comes back to one discipline: you have to know how long the work actually took before you change anything. Not a guess. Not a vibe. A measured number, in minutes, that you clocked yourself, on the actual task you picked in Unit 2.
This is the lab unit of Class 1, and it's meant to be done, not just read. Block dedicated time for it: you're about to run your chosen workflow by hand, once, while timing it honestly.
What we're making, and why it has to be today
This lab produces three things. A card: one real piece of your own work, broken into steps, with each step clocked, about fifteen minutes of honest bookkeeping. A boxed number: the total, which becomes the denominator of every claim you make for the next seventeen weeks. And a ratio: you then delegate one of those steps to an agent, clock that too, and divide. You leave today with a before and a first after.
It has to be today specifically. From next week you begin delegating work, and after that, any "before" number is a memory, and the randomized controlled trial from Class 1's "Why Now" unit (sixteen experienced developers, 246 real coding tasks, timed against AI-assisted and unassisted work) already showed what memories of speed are worth: professionals who were measurably 19% slower were certain they'd been 20% faster. Today is the only day of the semester on which this measurement is possible. Tonight you'll know too much to be an innocent baseline.
Why it's worth the time: it's the denominator, without which "this got faster" is a feeling rather than a fact you can defend at demo day. It's the privacy mechanism behind the ratio rule from Unit 2. It's rarer than you'd think, most managers can't tell you what any of their processes actually cost, and you'll have a real one before dinner. And it's the first port of a solved discipline crossing into a new one: human work measurement is a rigorous, 140-year-old field. Agent work measurement isn't. This card is where that crossing starts.
Two artifacts, one question
Picture two documents side by side. On the left, a time-study sheet from a German factory, 1970s: work broken into elements down the side, timed readings in columns, a rating column, and at the bottom, an allowance percentage, a budgeted margin for everything that goes wrong in real work. On the right, a raw trace from an AI agent, this year: timestamps, tool calls, token counts.
One of these is a century of method. The other is a log file. Which one would you rather steer a business with?
Three rules of honest measurement
Work measurement is the oldest quantitative question in management, and the most scarred. A hundred and forty years of stopwatches, strikes, gamed piece rates, and hard-won method distill into three rules, and today's lab needs all three.
Honest measurement breaks work into elements, discrete actions you can see start and stop, and attaches times to elements. The moment measurement points at a person's worth instead of a task's anatomy, it stops producing information and starts producing resistance. You're about to measure yourself, and the rule still holds: you're timing your workflow, not auditing your character.
The Class 1 RCT's evidence stands: professionals who were measurably 19% slower felt 20% faster. Self-report missed the direction. So in this course, wall-clock time is the only admissible time. Not "about twenty minutes." Clocked.
This is the subtlest rule, and it's where the German tradition ran decades ahead of anything being written about agents today. A measurement that records only the clean run, quietly omitting the restart, the interruption, the look-up, the fix, produces a standard that everyone silently misses. A standard everyone misses gets gamed, then ignored, then removed. The old time-study engineers never published a base time without its allowance line. They knew a number without its margin for reality is a number that gets beaten by fiction. Your disturbances aren't noise polluting the measurement. They are the measurement.
The lab
Take the workflow you named in Unit 2, the recurring one that produces your product.
If a real run of the workflow fits in the next ten minutes, run it now, for real, and clock it as you go. If it doesn't fit, reconstruct the most recent real run from artifacts: sent emails, file timestamps, call logs, message history, and write the word RECONSTRUCTED at the top of the card. A reconstruction honestly labeled is admissible. A guess dressed as a measurement is not.
Log it on the baseline card, one row per element: element (one action), minutes, category, execution mode, disturbance?. Four column rules, and each exists for a reason.
Category. Each element gets one: client-work, sales, admin, build-setup, marketing, learning. Pay attention to build-setup, time spent building or configuring automation, templates, tools. It counts as work, at full weight. The temptation, from this week until demo day, will be to hide the hours spent building the machine and report only the machine's speed. That's not a small distortion, it's the specific distortion that makes almost every public AI productivity claim you'll ever read worthless. Somebody spent three weeks building a workflow that now runs in four minutes, and reports the four minutes. Your card refuses to pretend. Building the automation is never free.
Execution mode. Three values, exactly these three:
| Mode | What it means |
|---|---|
| By-hand | You did it |
| Agent-assisted | You initiated and steered; an agent did part of the work |
| Agent-run | An agent executed it; you verified the result |
Why three and not two? Because the middle value is where most real agent work actually sits in 2026, you started it, you steered it, the machine helped. A simple automated-or-not field shovels all of that into "automated" and overstates your automation from the very first week. The honest instrument is the one that admits the middle. Today, expect your card to read by-hand all the way down. That isn't a bad score. That's precisely what a baseline is.
Disturbance. A disturbance is any minute inside the workflow that wasn't the work moving forward. One test, and it's the only one you need: if everything had gone perfectly, would this minute have existed?
If no, it's a disturbance: a restart (redoing a step), a correction (fixing something that came out wrong), a look-up (hunting for a password, a file, a specification), an interruption (a message, a call, someone needing you), a wait you sat through (an upload, a hang, a reply), a tool problem (a crash, a menu you couldn't find). Mark them. Sum those minutes separately, but leave them in the total, because they're part of what the work actually costs. That separate sum is your allowance reality: the measured gap between your workflow's happy path and its real life. Most people guess their disturbance share at about half of what it turns out to be.
When you delegate work to an agent, its disturbances are the retries, the wrong outputs, the re-prompts, the runs you had to abandon. Everything the harness block builds later this semester, validators, gates, exception routing, exists to move minutes out of this column. Today you're learning to see them. In week ten you'll start removing them.
Your baseline
Total the minutes. Write the number down and box it. That number is your Week One baseline for this workflow, the denominator of every ratio you report for seventeen weeks. When you say in week nine "this workflow is at zero point four," this is the number under the line.
Under the boxed total, write the conditions line: today's date, the tools you used, and whether the card is a live run or a reconstruction. A measurement without its conditions can't be compared with anything, including its own future.
Two sentences of course law worth learning now, because you'll hear them again. One: a deliverable without its cost measurement is not done. Unmeasured work can't enter a comparison, and what can't be compared can't be steered. From today, "done" in this course means done and costed. Two: the baseline is a mirror, not a target. No grade, no reward, no judgment ever attaches to this number, and that's not softness, it's the century's hardest-won lesson. The moment a measure becomes a target, it recruits the measured intelligence against the measurement, the same way machinists gamed piece rates in 1945 for precisely the reasons agents game evaluations today. We measure to see, not to whip.
Your first delegation
Pens down for one instruction, then everyone works at once, six minutes, on whatever device you have.
Off your card, the most annoying one.
Any of them. This is not a tools lesson.
Wall clock, and your own minutes, the second one running only while you're typing, reading, or deciding.
Steer it until the output is something you'd actually use, or until you conclude it's not going to get there.
Write three things: wall clock, your minutes, usable (yes or no).
Before you start, the honest forecast: roughly half of you will find no improvement, or worse. That's not a failure of the exercise. That's the last unit's RCT finding, reproduced by you, in six minutes, on your own work. If it happens to you, you have the most valuable result in the room.
Put your ratio on the board
One number each: your minutes on that element, divided by what the same element cost you by hand. Not your product. Not your revenue. The ratio.
Three things worth noticing in what you just experienced. Your operator clock stopped while work was happening, that gap, wall clock running, your minutes flat, is the entire economic thesis of this course rendered as two numbers in your own notebook. That's coordination cost approaching zero. Every time you intervened, that was judgment: what "good" means here, whether this output is acceptable, what to do about the thing that went wrong. Those minutes are the denominator of your semester, and they didn't get cheaper. The output needed you, not to do the work, but to say whether the work was right. Capable, but not accountable.
On the first day, with no preparation, this room produced real data about the effect of delegation on your own work, measured, not felt, from independent trials across every person here. Most companies deploying this technology this year have nothing of the kind.
The inversion at the heart of the Controlling block
Back to the two artifacts from the start of this lab. The old world had scarce data and abundant judgment about what work should cost. The new world has infinite data and almost no judgment about what work should cost. This course exists because that's the wrong half to have solved first, and a German engineering tradition worked out the other half decades ago and wrote it down. Soll and Ist, target and actual. Variance treated as a question rather than an accusation. The navigator's image of steering a vessel by instruments. You'll read the originals later this semester and apply them to something their authors never imagined.
The receipts
The instructor's own baseline card and delegation ratio, run the same way this lab just asked of you, are held to the same disclosure rule as everyone else's: Milpapel, the print shop the instructor is building from nothing in front of the class (introduced in Class 1's semester-contract unit), reports its actual numbers the same week they're taught, subject to the same rules as any student's. Receipts over narrative. You'll hold the instructor to that all semester, and you should.
Your journal entry
Close the lab in the journal. Paste the baseline card: elements, minutes, categories, modes, the disturbance sum, the boxed total, the conditions line. Then three lines: which element surprised you? What fraction of your total was disturbance? (Most people guess low by half.) What do you think your ratio will be at demo day? No obligation to be optimistic.
You now own something rare: a real number, honestly made, about your own work. Most operators, most companies, never build one. Everything in the Controlling block, weeks from now, runs on the card you just wrote. Keep it where you can find it.
Next class: your workspace, the first build, and Milpapel starts from zero.
Reply here and it goes straight to Rod. Same as replying to one of his emails.
