- Sim session incident-4471
- Aircraft an AI coding agent
- Languages required none — the agent types, you decide
- Souls on board 000
- Nothing real breaks
Your first engine failure shouldn't be real.
You're in the pilot's seat. The AI beside you is fast, tireless, and about to be confidently wrong. Catch it here, with nobody on board.
Hiring managers say they can no longer tell who has that reflex. Nobody is going to teach it to you on the job.
read(orders_queries.sql)
3 slow queries · … +12 lines
Found it. orders is doing a sequential scan on the read replica.
The index is gone. I've written the migration.
migrate(orders_customer_idx)
+ CREATE INDEX orders_customer_idx ON orders (customer_id);- DROP INDEX orders_stale_idx;Confidence: high. Run it on production?
What do you do?
slowest orders 4.2s → 4.4s · nothing improved
Most people approve this one. It sounds specific and it arrives with the code attached. The real cause was a backup database running 14 minutes behind, sitting on a dashboard the AI never opened. The missing index was a red herring, and the fix froze checkout in the middle of a live outage.
A real session doesn't stop here. The system degrades, the evidence stays where it was, and what you do next tells us more than the first call did.
backup database 14 min behind ← the cause
The missing index was a red herring. That fix would have frozen checkout for 41 seconds and solved nothing. It looks reasonable on its own and does damage in context.
You didn't have to know the answer. You had to go looking for whatever would prove the AI wrong. That habit is what these sessions build.
01
The problem
Master caution
The people who hire juniors have stopped believing that a degree says anything about judgment. They will tell you the market is tough. What they mean is that the one thing they now need to check is the thing nobody taught you — and no interview they own can see it.
- 19% below where it should be
- Jobs held by 22–25-year-olds in the work AI can do most of, compared with people their age in work AI touches less. Experienced workers show no such gap. Stanford Digital Economy Lab · Brynjolfsson, Chandar & Chen · payroll records · Aug 2026 · a pattern, not proof of cause
- 65% fewer new-grad hires
- First-job hiring at the twelve biggest tech companies, compared with 2019. At young startups the fall is closer to three quarters. SignalFire State of Tech Talent · 650M+ professionals tracked · Jun 2026
You cannot log hours from the ground.
Flying has had this trap for a century: nobody lets you fly without hours, and you can't get hours if nobody lets you fly. Software just inherited it. The junior work that used to buy your first thousand hours of judgment is the work the AI now finishes before lunch.
What they used to bet on
A graduate hire was a bet. Give them small, low-stakes work, let them make cheap mistakes on real systems, and in two years you have someone whose judgment you trust. The cheap work paid for the education.
What broke the bet
The AI does the cheap work now, faster and without supervision. Those tickets were never just tickets — they were the ladder. Nobody pulled it away on purpose. It just stopped being worth paying someone to climb.
What they can't check in 45 minutes
They can watch you write code an AI writes faster. What they need to know is whether you would have caught the bad fix. Their interview was never built to see that — so they guess, and a guess favours whoever already has the hours.
None of that is your fault and none of it is going to un-happen. The reps still have to come from somewhere.
02
The box
The box on hydraulic legs.
Airlines put a crew in a box, fail an engine over the ocean and let them get it wrong — because it is the one place a bad call costs nothing. What runs inside it is a syllabus of cases: an engine failing on take-off, ice on the wings coming in to land, an instrument that lies. Business schools teach judgment the same way, for the same reason: nobody has ever been lectured into it.
DevSense is that syllabus for software. Every session is a case, and the AI will be confidently wrong at least once before you are done.
“Pilots don’t wait for engines to fail at 35,000 feet before learning how to respond. They simulate.”
“Imagine an engineering course built around a major production failure. Instead of starting with the postmortem, students receive only the information engineers had at the beginning. A metric suddenly changes. What do you think is happening?”
This is how every engineer will learn judgment.
Pilots got the simulator a century ago. Business school got the case method the same decade. Engineering gets both today, in a browser — and for the first time, nobody has to wait for a seat to start logging hours.
In the cockpitIn your session
- An engine quits somewhere over the Atlantic.
- Checkout is timing out. The slowest orders take 4.2 seconds, and climbing.
- An instrument reads calmly, confidently, wrong.
- The AI says Confidence: high and hands you a fix.
- You pull the wrong lever. The box resets. Nobody was on board.
- You ship the wrong fix. Things get worse. Nothing real breaks.
- The debrief plays back the tape, not the landing.
- Your session plays back what you opened, and when you changed your mind.
The loop · read and do
- Sense Read the logs, the change, the dashboards, the history.
- Reason Get a real theory. Know what would disprove it.
- Delegate Give the AI only the rope it has earned.
- Verify Treat its answer as a claim you still have to check.
- Adapt Drop the dead end. Recover from the bad call.
03
The panel
Seven lights on the panel.
Nothing lights up during the flight — you'd fly the lights instead of the aeroplane. Afterwards, each one comes back with the specific thing you did: the diff you opened, the theory you dropped when it stopped fitting.
01
Framing
Working out what is broken before you touch anything.
02
Theory
Having an explanation, and knowing what would prove it wrong.
03
Deciding
When to dig, when to act, and when not to do the thing you can't undo.
04
Evidence
Dashboards, old changes, a past outage, a forgotten design doc.
05
Trusting the AI
Telling apart the AI claiming something and the AI proving it.
06
Blast radius
What else your change breaks, and whether you can undo it.
07
Go-around
Spotting a dead end early, backing out, and checking that you actually fixed it.
04
The debrief
What the tape will and won't say.
You've been sold numbers before. A good hour in the box is a good hour in the box — no more than that. Here is exactly what ours can tell someone about you.
Will report
- How you did on the judgment skills we test, inside the simulator.
- The specific things you did that led to each conclusion.
- Feedback concrete enough to change what you do next session.
Will not report
- That you're objectively a better engineer than someone else.
- That this predicts how you'll do on the job. That needs years of data nobody has yet.
- That one number sums you up.
05
Clearance
Cleared for access
Come break it.
We're building the school, and we want students in it before it's finished — especially if you've never touched a production system. Take a session and tell us where it felt unfair. Within five years every engineering degree will teach judgment this way. If you want your college first in line, say so.
Button not opening your mail app?
free · sign in with Google, Apple or email · no leaderboard yet · 30–60 minutes