🔍 Read the full analysis: Why Gewerkton Turned To AI Agents For Its Brand Relaunch on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get monitors, keyboards and dev gear delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Gewerkton, a construction-site documentation tool built in Iffeldorf, Bavaria, says AI coding agents implemented a broad HORIZON redesign on Oct. 2 and 3, 2026, spanning its web and phone apps, Studio, 26 icons and a website with 529 pages in 40 languages. Founder Thorsten Meyer approved the written plan and released each part after review, according to the account published by the company’s publisher, which discloses that it builds Gewerkton. The project offers a case study in using AI agents for a real product launch while keeping consequential decisions and final checks with a person, as detailed in the original analysis.

The relaunch updated the web app’s design system, moving 281 hard-coded colour values to design tokens while retaining 11 exceptions, and restyled 47 cards. Gewerkton says the phone app’s version 0.5.0 adopted the new look on iPhone and Android, and its Studio gained light and dark versions. All 26 icons were re-rendered in Blender in cobalt and navy. The website’s opening section was rebuilt around product screens, and the company reports uploading 1,583 files without an error.

The work also included 228 AI film runs for five regions, of which 222 had region-specific versions, plus eight help videos. The company qualifies its description of the work as taking “two days”: an interim design draft dated Sept. 26 was already available, so the timeframe refers to execution from an approved plan through the live website, not the entire design process.

Meyer directed a fleet of Codex and Claude coding agents, which worked in a separate code copy and branch—an approach also explored in Claude’s agent teams. The plan listed open decisions with options and recommendations; Meyer approved the recommendations. According to the account, no work was merged or delivered until he released that part. Films showing the product waited for his review of the new interface, and paid film reruns also required his explicit approval. Apple’s developer agreement presented a separate human-only gate: Meyer had to accept the latest Program License Agreement before the iPhone build could proceed.

Human Gates in an Agent-Led Relaunch

The project’s reported process illustrates a distinction between delegating implementation and handing over product authority. Agents carried out planned changes, but a person approved the plan, checked results and controlled releases. That arrangement matters to teams weighing whether coding agents can handle coordinated updates across a product, especially when a change touches interfaces, localized content and paid media.

Gewerkton’s account also shows why passing automated tests is not the same as confirming a release is ready. A human review caught white icon tiles that obscured nearby labels in dark mode, an issue no test had flagged. Visual inspection of film outputs found missed red details and other recolouring defects. The examples are the company’s own project report, not independent evidence that the same workflow will produce similar results elsewhere.

How the Release Was Checked

Gewerkton says its release checks combined automated tests, multiple screen sizes and languages, and visual review. The web app was inspected in 200 views across English, German, Japanese and Hindi, at widths of 390, 834 and 1,440 pixels, in light and dark modes. The company reports no overflow or console errors in those checks. Its phone app passed 853 of 853 tests, the Studio passed 2,158 of 2,158, and the website passed seven test suites, followed by browser checks across three widths and five regions.

Some web-app tests still failed: eight tests across seven files. Gewerkton says six files also fail on the main branch and one test depends on execution order and passes when run alone. On that basis, the company says the redesign did not cause those failures, but also did not resolve them. Film checks were more manual: reviewers inspected the middle frame of 157 mirrored films and watched reworked films again at 10%, 50% and 90% of their duration.

The visual review led to several corrections. The icon issue was fixed with transparent treatment for pale areas, a separate dark variant and a layout adjustment. For films, an automatic mask intended to remove red brand elements missed red in shadows and warm lighting. The company says 24 files were cleaned and 10 regional versions with leftover red temporarily switched to a European cut; nine runs were later reshot with Meyer’s approval and passed visual checks. A recolouring attempt that turned a person’s lips blue was discarded.

“Gewerkton is built by our publisher — we build it ourselves and write down what we learn.”

— Gewerkton’s publisher, in its disclosure

Limits of the Project Record

The account is a publisher-provided description; it does not include independent verification of the reported tests, file counts, agent contributions or production costs. It also does not provide a full breakdown of human labor, the time spent preparing the interim design, or a comparison with a non-agent workflow. The reported scope and results should be read as Gewerkton’s account of its own relaunch, rather than a general measure of how reliably AI agents can deliver similar work.

The supplied material says the film project has one recorded cost figure, “about 134 dollars actual,” but ends before specifying what that amount covers. It is not clear whether this is the full cost of the film work, a subset of expenses, or how it compares with other production approaches. The record also does not say whether every regional film was ultimately replaced after the temporary use of European cuts.

For the Android app, Gewerkton reports five major findings from testing on a Pixel 10a, none considered a blocker. It identifies a button label breaking mid-word because the new design uses larger text and more padding, but the supplied account does not describe all five findings or confirm that any fixes have shipped.

Planned Android Follow-Up

Gewerkton says all five major Android findings are scheduled for version 0.5.1. The supplied material does not give a release date, so it remains unclear when users can expect that update. The company’s account also says nine film runs were reshot after visual review; it does not set out a further publication schedule for any remaining regional versions.

The documented process leaves future changes dependent on the same release gates described for this relaunch: agent implementation in a separate branch, checks on the result, and human approval before release. Whether Gewerkton will use the approach again, or publish further results on time and cost, has not been stated.

Key Questions

What did Gewerkton relaunch?

Gewerkton says the HORIZON redesign covered its web app, iPhone and Android apps, Studio, 26 icons and a 529-page website available in 40 languages. The work also included regional promotional films and help videos.

What role did the AI agents play?

According to Gewerkton’s account, Codex and Claude agents carried out implementation from a written plan in a separate code branch. Founder Thorsten Meyer approved the plan, reviewed outputs and authorized each release.

Did the relaunch take only two days?

The company says the execution ran on Oct. 2–3, 2026, from an approved plan to the live website. An interim design draft from Sept. 26 was already in hand, so the two-day figure does not cover the full design process.

What problems did testing uncover?

Visual review caught icon tiles that obscured labels in dark mode and film recolouring that missed some red details. Gewerkton also reports five major findings in an Android test, none a blocker, and says they are scheduled for version 0.5.1.

Is the reported cost of the film work clear?

No. The supplied account gives a figure of about $134 in actual cost but ends before explaining what it covers. It does not provide a complete cost breakdown or a comparison with other production methods.

Source: ThorstenMeyerAI.com