What changed
Give it eyes. Then take everything off the human's desk that does not need a human.
My day job is keeping a test suite of 1,600 automated tests running against a large product that ships features fast. When the product changes, tests break, and someone works out why. The AI was good at the repair and blind to everything else, so I was the courier.
Before
One broken test. Ten hand-offs.
- Me, reading the failure on a dashboard
- Me, copying the error into a chat
- It guesses at a cause it cannot see
- Me, screenshotting so it can see
- It writes a fix
- Me, pasting the fix and running it
- Me, waiting, then copying the next error
- Repeat until it passes, or the day ends
After
One broken test. Two hand-offs.
- It reads the failure and the video itself
- It opens the product and confirms what changed
- It writes the fix and runs it twice to prove it
- It lands on a board, marked ready, with the reason
- Me, pasting and running
- Me, confirming it passed
Porting it
The same three jobs get automated every time. The same one stays yours.
Whatever the work is, the AI gathers the evidence, does the work and proves it. You keep the one step that is hard to undo. Only the words change.
I read two items a day instead of thirty failures, and I never open a log again.
Every PR reaches you already cleaned up, with the tests green. You spend your review on the design call, not on naming and formatting.
Nothing ships on vibes. A prompt cannot be promoted until the scores are recorded against that exact version, and regressions surface before you look.
Tickets land already reproduced with a proven fix attached. The ones that turn out to be user error never reach an engineer at all.
You walk into each conversation already briefed, and the admin trail chases itself. No screening decision is ever made for you.
One screen that tells me what genuinely needs me today, and hides everything that does not. Same machinery, no testing in sight.
Every row above has the human gate in a different place, because that is a choice, not a limitation. Move it earlier if the work is risky, later once you trust it. I am building the personal one next.
The part that makes it trustworthy
The AI cannot mark its own homework
An AI will tell you the work is ready, because it believes that. So "ready" here is not an opinion. It is a gate, and the evidence has to already exist.
- Two clean runs recorded on my machine
- One clean run on the real test machines
- All three against the same version of the code
- The run id written onto the card, so I can check it
The part that compounds
It files its own bugs and learns from its own mistakes
The system is not finished and was never meant to be. It gets a little better every day, and most of that improvement is not mine.
A tracker that fills itself
I can file a bug or an idea with one key. The interesting half is that the AI files them too, on its own, from whatever it ran into that day: a rough edge in the tooling, a gap it noticed, a bug in my own scripts.
Every row says who raised it, so I always know which are mine and which it brought to me. I keep, reprioritise or bin them. Nothing gets built without me saying so.
◉ TRACKER 6 open ──────────────────────────────────────── 🐛 BUGS 👤 Newest card sorts to the bottom high 🤖 Demo page could leak real names high 🔧 CHANGES 👤 Cap every panel at four rows med ✨ FEATURES 🤖 Shareable demo mode for the board med 🤖 Re-queue runs that stall silently low 👤 Live "what am I on right now" high ──────────────────────────────────────── 🤖 raised by the AI 👤 raised by me
It writes the lesson down
If I correct it, or it repeats a mistake, the rule goes into a permanent file that same day. Not a note for later. Part of how it works from then on.
It runs its own retro
It re-reads the day, finds what it got wrong, and writes down anything the live capture missed. It also scores its own past predictions to see where it was overconfident.
It adds a test against itself
A lesson becomes an automated check, and the check has to fail on purpose first. If it cannot catch a planted mistake, it is not trusted.
The practical effect is that the same problem does not come back, and everything it learns about the product is written down where the next investigation will find it. The bigger the suite gets, the more that matters.
The four screens
Each one answers exactly one question
They sit open on a second monitor. The AI writes to them as it works. Real panels, names changed.
What needs me right now?
Headquarters
The only screen I have to look at. Everything else exists to keep this list short. On a good day it holds two items.
◉ HQ 9:42 PM ──────────────────────────────────────── 👉 YOUR MOVE (2) 1 Paste 3 files, then run acc-07 2 Re-run these 9. No fix needed ▶ RUNNING (1) bil-19 local run 3m 21s elapsed ✉ MESSAGES (0) nothing waiting 📌 TRACKER 3 bugs · 2 changes · 5 ideas
Where does every broken test stand?
The Mission Board
One card per test, ordered by how close it is to done. A card that has not moved in three hours turns red on its own.
◉ MISSION BOARD 31 in this batch ──────────────────────────────────────── ♻ PASSED ON RE-RUN just try it again ♻ chk-04 23m ago 🛰 PROVEN ON THE REAL MACHINES 🛰 acc-07 41m ago 🔧 FIX READY waiting on you 🔧 onb-05 31m ago why: the Continue button moved 🟡 STILL INVESTIGATING 🟡 adm-03 1h 14m ago
What is running, and did it pass?
Flow Ops
Several tests run at once, in their own lanes. Before, they went one at a time, and most of the day was waiting.
◉ FLOW OPS 4 lanes open ──────────────────────────────────────── ▶ RUNNING (2) bil-19 3m 21s ████████░░░░ srch-02 1m 04s ███░░░░░░░░░ 🏁 RESULTS newest first ✓ acc-07 passed 4m 02s ✓ log-06 passed 3m 48s ✗ adm-03 failed 2m 11s
What actually got done today?
The Workshop
The day's record, and arrow keys back through previous days. Every number on this page came off this screen.
◉ WORKSHOP Tuesday, Sep 1 ──────────────────────────────────────── 🏆 37 tests fixed and signed off ⏱ fastest today: 6 minutes ✅ WINS 9:14p crt-14 signed off 12m 8:52p prf-05 signed off 28m 8:30p chk-02 signed off 6m … 🐛 1 sent back as a product bug
One real day
September 3rd, start to finish
Not a demonstration. One ordinary day, taken off the Workshop screen.