Towards the Dark Factory: Where the Human Still Is

Part 5 of 5 in the “Towards the Dark Factory” series
Part 1 opened on a wager: can you adopt a language you do not know by running the process rather than writing the code? Four parts later, Sestava is a native Swift app on TestFlight, built by agents behind a wall, setting their own pace and writing their own backlog. This last part is about the person. Not in the abstract: the record of six days says exactly what the owner did, and it sorts into five kinds of work. Then what it would take to remove the merge button, which is Saumya Tyagi’s level 3.5, and what level 4 would need from here. Then the wager.
1. Rulings on product facts
The most frequent thing the owner does is answer a question the code cannot. Three from 29 and 30 September, because each one taught the process something.
Calories. The owner’s device pass found a Summary that read “Workout, energy, heart rate” with energy and heart rate empty. Asked, he ruled: the app does not estimate calories for a phone-only workout; calories and heart rate are shown only when a watch measured them. Eighteen minutes later the ruling was found to contradict a spec merged that same morning, which chose a net MET estimate for iPhone-only sessions, shown as “Estimated”. The coordinator had recommended without reading the spec, and the owner had agreed. The ruling is suspended on the issue; nothing was implemented from it; the rule that a recommendation touching a merged spec quotes the spec’s section came from it. A person’s ruling is an input to the process, not an override of its own records.
The watch at launch. Sestava Pro was to gate mirroring to a watch. On 29 September the owner ruled that without the watches Pro makes little sense on launch day, and brought the Apple Watch and Garmin fēnix apps forward on the board, while Garmin Connect and TrainingPeaks sync are hidden everywhere until they exist. A day later the launch moved to 17 November to make room for the Apple Watch app, which is planned for launch and not yet shipped as this is written, and the Garmin app moved to the first update after launch. That one ruling brought two issues forward on the board, opened one to hide the sync everywhere, and changed what the App Store listing may claim.
The privacy page. A writer checking the first blog post against the code found that the privacy page named body weight and date of birth as Health reads the app does not make, and an Apple Watch that did not exist, and filed a bug. The owner’s ruling: the privacy page describes the app at launch, not today’s TestFlight build. The page keeps all three, and the fix is one-directional: any Health type the code reads or writes today that the page does not name. That found one, resting heart rate, missing in both languages, and a test now pins every read and write type in the Swift source to both pages. The rule: a copy check states what the code does today and what the merged specs and the owner’s rulings say ships at launch; a claim that is false today and true at launch is a dated claim, not a bug.
As of 30 September, 18 issues carry the needs-petr label, and the 29 September retrospective notes three pull requests that once waited on 19 decisions. The mitigation from Part 3 is provisional rulings for technical questions; product facts still wait.
2. Reading the Czech
Sestava ships in English and Czech, and the agents write both. What they cannot do is judge the Czech: the on-device model does not support Czech and Apple’s translation framework has no Czech (as measured on the development Mac), and the models we use write it well enough to be dangerous. So the process treats Czech as a claim that needs a native reader. A wave that typed every error sentence in the app kept its 117 Czech drafts in a table outside the string catalogue, marked for review, with a UI test pinning that a Czech device sees English until they are read. Its pull request says near the top: the owner reads the five Czech lines on the format page before merge.
The same is true of the website and this blog. The Czech site’s plan carried eight rulings for the owner (formal or informal address, the price in crowns, whether a Slovak reader gets Czech, whether Apple’s Czech names are right); as of 30 September four Czech blog drafts waited for his read (he had read and published the first that day); the glossary that fixes terms lives in the repo, and a term changes there first. The rule that came out of one bad afternoon: a translation reports a doubtful claim and never repeats it silently.
3. Merging pull requests
The owner merges every pull request, and the merge button is the ship button. Xcode Cloud archives the app from every push to main and hands it to TestFlight; GitHub Actions is the only test gate. So a merge is a release to the internal testers, and the owner’s review of a pull request is a review of claims: the body ends with a table of gates, each naming the commit it ran on and its result, and a device checklist for him. He does not read the Swift. He reads whether every gate ran after the last commit, which is a rule because once it had not.
This is Tyagi’s level 3 exactly, “you approve the merge”, and it is the one step in the process that is human by design rather than by limitation.
4. Holding the phone
The simulator has no haptics, no Taptic Engine, no pocket. When the step-change tick turned out to be unnoticeable while walking, no test could have known. The fix shipped with a 24-line device checklist whose instruction is “for every line, write down what was felt, not what was expected”: in the hand on screen, in a trouser pocket with the screen on, locked in a pocket ten minutes in, Control Centre pulled down at a step, an incoming call, AirPods connected mid-workout, Low Power Mode, the phone face down on a gym floor. The launch fix from Part 3 has its own list: cold start then start a workout at once, cold start from the widget, reminders after a relaunch, lock the phone immediately.
The owner’s device pass on build 2 produced ten findings in under an hour one morning, and one of them (the voice and the Lock Screen banner out of step) is still a hypothesis waiting for device logs. Device facts cost a TestFlight build each. The optimisation considered is a debug screen in the app for such checks; it is not taken yet, and the checklist in each pull request is the current answer.
5. Accounts and signatures
The last kind is the least interesting and the least delegable: Xcode Cloud was created from the owner’s Xcode, the in-app purchase was created in App Store Connect with his yes (a non-consumable, with Family Sharing on, which cannot be turned off again), the developer team, the domain, the sign-ins. The retrospective’s advice is to batch these into one session with the owner present rather than pay a round-trip each.
What auto-merge would need here
Tyagi’s level 3.5 is “Same as 3, except some services auto-merge without you.” Could any of Sestava’s pull requests? The honest answer is: a class of them, with three things added.
A classifier for the device surface. A pull request that touches haptics, audio, the Live Activity, Czech copy or anything the privacy page describes needs a person, because the wall cannot see those. A pull request that touches only the core package (a Foundation-only Swift package whose gates are 100% coverage, the mutation runner and the schema check) or only docs does not, in principle. Docs-only and site-only pull requests are already treated as cheap; they would be the first to go.
The full mutation gate before the merge, not after. On pull requests CI runs the runner over changed files only; the full gate runs on main and weekly. That is fine while a person merges one wave at a time and a full gate runs locally before the push. For auto-merge the full gate would have to run on the pull request, which is hours of CI per merge; the runner’s timeout following the measured test time is an open issue that would make it shorter.
Claims pinned by tests. The privacy page now has a test that fails when the code reads a Health type the page does not name. The site’s feature claims do not, and the blog’s do not. Until a claim about the product fails a build the way a line of code does, a person reads the copy, and copy travels in the same pull requests as code.
None of that is built. We are saying what it would take, not that we are there.
The road to level 4
Level 4 is specs in, merged code out. From here, three gaps, in the order we would close them. The ticket-to-spec step still runs through a brainstorming session with the owner and an adversarial review; the review could run without him, the brainstorming, for a product, should not. The device surface needs either a debug screen that turns a felt property into a logged one, or the acceptance that some pull requests are never level 4. And the Czech needs a native reader until a model reads it as one. Everything else in Tyagi’s list, the wall, the backlog wiring, the weekly clean-up (here, the weekly full gates and the grouped dependency updates), exists here in some form already.
The wager, settled
The wager was whether a language you do not know can be adopted by running the process. Here is what the record shows, and then what we think it means.
Between the evening of 24 September and 17:39 on the 30th, 726 commits landed on main, 82 pull requests were merged, and two Swift packages and an app grew to about 18,400 lines of source and 22,200 lines of tests. 575 of the 582 non-merge commits carry an agent’s co-author trailer. The owner’s account is that he wrote none of the code; the agents commit under his name, so the log can neither prove nor refute it, and he has confirmed it. His work in that week was the five kinds above: rulings, Czech, merges, the phone, accounts. He read specs and plans before approving them, in prose; he read pull request bodies, which are claims and tables; he never had to read a Swift diff, and as far as the record shows he did not.
So: as a bet on shipping in a language you do not know, the wager is won, and won faster than either of us expected on the evening of the 24th. A native iPhone app with a workout engine, an open format, widgets, a Live Activity and a web editor reached TestFlight in five days, and the wall from Part 2 is a stronger verification than our older products have.
As a bet on knowing the language, it is lost, or at least not what the words suggested, and we would rather say so than let the first half stand for the whole. The owner does not know Swift now in the sense of being able to write it. What he adopted is the platform: Xcode Cloud, TestFlight, the simulator’s limits, HealthKit’s permission model, App Store Connect’s irreversible switches, what a Live Activity can and cannot update. And what he learned to do is direct Swift work and judge it by its claims and its gates, which is the work of a lead, not of a programmer. The language lives in the contract file, the specs and twenty-two thousand lines of tests, and that is where it will stay.
We think that is the correct outcome and the useful thing to know. The process did not make a Swift programmer; it made a Swift product, with a person in the five places the record says a person is still needed, and a written account of why. The launch date, decided by the owner on 30 September, is 17 November 2026; the pre-order is an open issue on the board as this is written. This part publishes two and a half weeks before it, and the owner has read and confirmed the sentences above about what he can and cannot do in Swift.
This is Part 5 of the Towards the Dark Factory series. Part 1: Level 2 to Level 3 in a Week. Part 2: The Wall. Part 3: The Agent Sets the Pace. Part 4: The Agent Feeds the Backlog.
Related: 140 Pull Requests in a Week · Architecture Tests as a Seatbelt Against Agentic Drift
Cover photo by Tasha Kostyuk on Unsplash.


