How long does it take to build an app with AI? Our 58-day timeline

The specification fixed the decisions the rest of the work depended on: iPhone first and adapted for iPad, offline first, and a permanent free tier. Before launch we cut the free tier from 50 items to 30. Before the 1.0 build was archived, the full automated test run passed 901 Swift package tests, 1,313 app unit tests and 148 UI tests.

Apple says App Review typically reviews at least 50% of submissions in less than 24 hours and 90% in less than 48 hours. Updates have gone through quickly: 1.8 was approved 5 hours 14 minutes after we submitted it, and 1.9 after 5 hours 39 minutes, of which 3 minutes were spent in review. We release by hand once a build is approved, so a release date can fall a few hours after approval. Approval is never guaranteed; none of Countstead's submissions has been rejected so far.

The agents

Up to the commit version 1.0 was built from, 391 of the 550 commits (not counting merges) name a Claude model as co-author, mostly Claude Fable 5, Claude Opus 5 and Claude Fable 5.1. By version 1.9, 1,789 of 2,229 commits do, and most of those since launch name Claude Opus 5.5. Each piece of work started as a numbered plan, a written brief an agent worked from: 292 of them before launch and 481 by version 1.9.

A person

A person decided what Countstead should be and what each release would include. That covered the specification, the free tier, pricing and the 1.0 scope. They tested builds on a real iPhone through TestFlight, working through a release checklist, and they set up App Store Connect. The agents cannot submit to App Review on their own: their permission check treats it as a real-world transaction and blocks it, so a person presses Submit or approves the call.

Even when agents write most of an app's code, human experience and intuition play a major part in whether it succeeds, and they far outweigh the coding. On Countstead they guided choices such as cutting the free tier from 50 items to 30 and settling the 1.0 scope.

We treat the agents as fast developers who still need a written brief and a reviewer.

Deciding what to build

The first commit was a specification, and every later plan referred back to it. The scope of version 1.0 was confirmed on day 30, more than three weeks before submission. An agent can implement a decision quickly, but someone still has to make it.

Tests and review

Each release build had to pass the full automated test run first. Reviews by separate agents caught real problems. In one October review, five reviewers raised about 90 findings on a set of plans, 11 of them blockers, before any code was written. In another, the first review round found two database operations that could deadlock each other, and the fix went in before the change merged. That test run became a bottleneck of its own: on 10 October it took 120 minutes, and by 13 October it was down to 29.4 minutes.

App Store setup

Two App Store Connect steps held up the 1.0 submission. "Add for Review" failed until we entered a privacy policy URL for each of the app's 13 App Store localizations. Apple lets you localize that URL, and the 12 localizations we had just added had none. App Store Connect refused the subscription group when we submitted it on its own, because the group has to go in with its subscriptions. Apple also requires the first subscription to be submitted with a new app version, so version 1.0, the group and both subscriptions went to review as one submission.

Waiting for review

Version 1.0 went to review on 4 September and reached the App Store on 10 September. Later updates were approved within hours. Plan for the first review to take longer than the updates that follow.

Andrej Karpathy coined the term in February 2025. He described accepting every change the AI made without reading the diffs, and called the approach "not too bad for throwaway weekend projects". We agree: it is a quick way to try an idea or build a tool only you will use.

Countstead was built with AI agents too, and their work was checked. Each piece of work started as a written plan, separate agents reviewed the plans and the changes, and every release build had to pass the full automated test run. One review round found two database operations that could deadlock each other, and the fix went in before the change merged. With vibe coding, nobody looks for problems like that.

Building with agents still takes a software expert: someone who makes the correct technical choices, can tell when an agent is wrong and steers it in the right direction, instead of blindly accepting everything it writes. The deadlock is an example: a person saw a problem in the app's telemetry, had the agents write a fix and asked for the fix to be reviewed, and that review caught it. Making the app offline first was another, fixed in the specification a day before the first app code was written.

Vibe coding does nothing for the parts of Countstead that were never code: the product decisions, testing on a real iPhone and the App Store Connect setup.

An app keeps changing after launch. Countstead had ten updates in its first seven weeks. Changing an app that often without breaking it takes someone who understands the code. Karpathy wrote that, working this way, "the code grows beyond my usual comprehension".

If your app will hold customers' data, take payments or go on the App Store, someone who knows what to check should read and test the code before it ships.

1. Write the brief before anyone writes code. Countstead's first commit was its specification. Our app requirements template covers the same ground.
2. Budget as much time for testing and review as for building. Agents produce code quickly, and every change still needs checking.
3. Keep a person on product decisions and on the release button. On Countstead, the tooling enforces the second.
4. Treat App Store setup as its own task: accounts, privacy details, subscriptions and screenshots. Start it well before the build is ready.
5. Ask any developer how they use AI and how they review its work. Our questions to ask an app developer cover testing, ownership and handoff.