I put your calendar, mail, health, reminders and notes on one board, with a morning brief written from what is there. Hold a card to move or resize it, undo one change at a time. GPT-6 Astra in Codex wrote most of the code. Numbers in the first comment.

How did Astra change the scope or ambition of what you built?

Before Astra I was one person shipping one lane at a time. Since August 17 each task has been a git worktree with a written brief; Astra runs inside it in the Codex CLI at low effort, and a Claude agent wraps the lane to run the simulator, the 4,235 test gate and the device install that Astra cannot reach from its sandbox. That is how the board edit mode, the notes rework and seven review fixes in one night got built. 200 of my 3,175 Codex sessions were Astra, 37 on the day we cut the review build. The bigger change was the shells, the 3D bodies the face lives in. I drove Astra in Blender myself: the Pocket master now has 15 mechanical controls wired to the native animation engine, down to a battery door that releases its latch before it slides, and Mk I got a fitted body with real drive and vent openings across all 12 skins. It is the best I have used inside Blender. When a Pengu upgrade did not hold up I reverted it. The app went live on the App Store this week.

I have been building Mitamaton, a chief of staff app for iPhone, since the spring. It went live on the App Store this week. Since August 17 a large share of the code has been written by GPT-6 Astra. I went through my Codex session logs to get a real count: 200 Astra sessions out of 3,175 total, most of them between September 5 and September 14. September 5 was the day we cut the App Review build, and Astra had 37 sessions that day. September 12 was the notes rework, and it had another 37. The way it works is not one big agent. Each task is its own lane. A lane is a git worktree with a written brief, and Astra runs inside it at low reasoning effort in the Codex CLI with workspace write access. Astra cannot reach the iOS simulator from its sandbox, so a Claude agent wraps each lane and runs the test gate, the UI tests, the screenshots and the device install. Then a different model reads the diff as an adversarial reviewer before anything lands. Right now the gate is 4,235 tests on the iOS target plus about 5,600 on the backend. One thing I want to be honest about is how often the result was a red, and how often Astra just said so. On September 5 I asked it to write an eval for how the chat refuses a request once a usage pool is empty. It wrote the eval and the test, ran them, got 8 of 10 passing, and reported FAIL with the two production defects traced to their lines. It could have marked them as expected failures to make the run look clean and it did not. A sibling lane fixed both. On September 12 a notes lane crashed the app on Save after typing into the new iOS 26 rich text editor. The crash report showed UIKit's proofreading shimmer calling a selector the SwiftUI text view does not implement on the beta runtime, with no test frames on the crashing thread. So that one was the OS, and the lane shipped the feature around it. Lanes bounce, and I want to say that plainly. One board lane went back twice, once for a backend registry drift and once for coupling to a flag that was off. Another lane overwrote a test that encoded a decision I had made. When I restored that test it caught a half pixel regression in where the face draws. So the rule we ended up with is that effort only goes up when a lane bounces twice on the same finding. Here is what Astra built that you can actually touch in the app. The board edit mode: a hold on a card is the only way in, and the corner grips you can see are the ones that work. Undo is a back arrow that walks back through every move and resize. The notes rework, with search, tags, folders, soft delete and the markdown body. Seven smaller lanes in one night for the review cut, including the light mode contrast fixes and the crash symbol upload after a TestFlight build. The part I did most directly was the shells, the 3D bodies the face lives in. I drove Astra in Blender myself rather than through a lane, and it is the best I have used inside Blender. The Pocket master now has 15 mechanical controls wired to the native animation engine, with a battery door that releases its latch before it lifts and slides. Mk I got a fitted body with real drive and vent openings across all 12 skins. A still render hid two motion defects that only showed up in simulator playback, and a Pengu upgrade did not hold up, so I reverted it. The app itself does not call Astra. The backend routes on other models and I am not planning to change that for a contest. What it changed for me is the amount one person can build in a week. If you want to know more about the setup, ask. The lane briefs and reports are plain markdown in the repo and I can paste one.

How did Astra change the scope or ambition of what you built?

Mitamaton

A chief of staff for your day, on one board.

A chief of staff for your day, on one board.