🔍 Read the full analysis: What Gewerkton’s Two-Day Relaunch Reveals About AI Agents on ThorstenMeyerAI.com

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

Get office and shipping supplies 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 by the publisher of ThorstenMeyerAI.com, replaced the look of its web app, iPhone and Android apps, Studio and multilingual website on Oct. 2-3, 2026. The publisher says AI coding agents implemented the redesign from written plans, while developer Thorsten Meyer approved the plans, reviewed the results and authorized each release. The original analysis offers a practical example of how agent-led software work can move quickly while still relying on human decisions and visual checks.

The new design, named HORIZON, was applied to the web app, phone app, Studio, 26 icons and a website with 529 pages in 40 languages. The publisher says the web app’s 281 hard-coded colour values were moved to design tokens, with 11 exceptions retained, and 47 cards were restyled. The phone app reached version 0.5.0 for iPhone and Android; the Studio received light and dark versions of the new look.

The website also received a new opening section featuring product screens. The publisher reports uploading 1,583 files without an error and producing 228 AI film runs for five regions, with 222 runs made in regional versions, as well as eight help videos. It qualifies the “two days” description: an interim design draft existed on Sept. 26, so the period covers execution from an approved plan, not work starting from a blank slate.

Work was done in a separate code copy on an agent branch. According to the account, Meyer released each part before it was merged or delivered. Films showing the redesigned product were held until he had reviewed and approved the interface on Oct. 3, and paid film reruns also required his explicit approval. An Apple Program License Agreement blocked the iPhone build until Meyer accepted it, a step the agents could not perform.

Human Checks Shaped the Agent Workflow

The account’s main finding is not simply that agents changed many files quickly. It describes a workflow in which agents handled implementation but people controlled plans, releases and paid reruns. Meyer approved a written plan that listed open decisions, possible options and a recommendation; deviations were recorded, with the conservative option chosen when the plan did not hold up.

That separation matters to teams considering coding agents for production work. A fast implementation can still produce defects or require decisions that depend on access, policy or product judgment. Here, Apple’s licensing step required a person, and product films were gated on human review. The publisher’s disclosure makes clear that “agent-led” did not mean unattended or fully autonomous.

The reported checks also show why passing tests is not the same as verifying a finished product. The web app was inspected across 200 views, four languages, three screen widths and light and dark modes. The phone app passed 853 of 853 tests and Studio 2,158 of 2,158, according to the publisher. Yet visual inspection found an icon problem that automated tests had missed. That distinction is relevant to software teams deciding how much human review to keep in an agent-assisted release process.

The figures and lessons come from Gewerkton’s own publisher, which disclosed that it built the product. They are a company account of one project, not an independent comparison of agent performance or evidence that the same pace and results would apply to other products.

A Redesign With Measured Release Gates

Gewerkton is described by its publisher as a construction-site documentation tool built in Iffeldorf, Bavaria, by Thorsten Meyer, who directs AI coding agents from Codex and Claude. For this relaunch, the publisher says the agents worked from written plans in a separate branch; Meyer reviewed plans and outputs and released each stage. The process began with a design draft dated Sept. 26, six days before the execution period began.

The design work included a contrast adjustment. The original cobalt colour measured 4.38:1 against the light background, below Gewerkton’s stated working standard of 4.5:1 under WCAG 2.1 AA. The team darkened it to 4.80:1 instead of shipping the original shade. The publisher reports that the website was checked in five regions at widths of 390, 834 and 1,440 pixels, and that the live site was tested again after upload.

Films required checks beyond code tests. The publisher says reviewers inspected the middle frame of each of 157 mirrored films and watched each reworked film again at 10%, 50% and 90% of its duration. Those steps followed problems with automated recolouring, including missed red details and attempts that changed lips, faces or stickers. The account says 24 files were cleaned and 10 regional versions with remaining red were temporarily replaced by the European cut; nine were later reshot with Meyer’s approval.

“The two days cover the execution, from the approved plan to the live website.”

— ThorstenMeyerAI.com, Gewerkton’s publisher

Test Gaps and Fixes Still Open

The publisher reported eight failed web-app tests across seven files. It said six of those files failed in the same way on the main branch, while another test depended on execution order and passed when run alone. The account says the redesign neither caused nor fixed those failures. It does not provide independent verification of the test results or a full accounting of how all remaining issues will be resolved.

Visual review also found defects that the automated checks did not flag. The first icon version had white tiles that obscured nearby labels in dark mode. The fix made white and pale grey areas transparent, added a separate dark variant and corrected a layout rule; the publisher says it went live that evening. Film recolouring also proved unreliable: red in shadows or warm backlight could escape the mask, while broader masks sometimes altered unrelated parts of an image.

Film production was paused after round nine, when the publisher says 38% of jobs were failing and 208 of 228 runs were complete. It reports that 76 failed jobs were not billed. The provided account ends just as it begins describing the film cost, giving only an approximate figure of $134 in actual costs; it does not specify the full cost basis, all work involved or a comparable budget. The account also does not establish whether the reported schedule, test coverage or cost can be reproduced on other projects.

An Android test on a Pixel 10a found five major issues, none classed as blockers by the publisher. One involved camera-screen buttons breaking mid-word after larger type and padding were introduced; the German cancel label wraps onto two lines. The company has scheduled all five findings for a later app release, but the supplied material does not give a release date.

Android 0.5.1 and Follow-Up

Gewerkton says the five Android test findings are scheduled for version 0.5.1. The publisher has not stated when that update will ship, so the timing and final resolution of the button-wrapping issue remain unconfirmed in the supplied account.

The immediate next step is to address those findings and continue checking the released redesign. The publisher says nine of the regional film runs that had been switched to the European cut were later reshot and visually reviewed. It has not said whether the remaining temporary substitutions will also be replaced, or provided a complete schedule for further film revisions.

For readers evaluating the project, the reported results should be read as a single company’s account of a particular release. The next useful evidence would include the timing and outcomes of the Android fixes, the disposition of outstanding test failures, and clearer cost details for the film work. Until then, the account supports a narrower conclusion: AI agents completed substantial implementation work, but human approvals and inspection remained part of the release process.

Key Questions

What changed in Gewerkton’s relaunch?

The HORIZON design was applied to the web app, iPhone and Android apps, Studio, 26 icons and a 529-page website available in 40 languages. The publisher also reports 228 AI film runs for five regions and eight help videos.

Did AI agents complete the relaunch without human involvement?

No. The publisher says AI coding agents implemented the approved plans, while Thorsten Meyer approved the plans, reviewed results and released each part. He also had to accept an Apple license agreement before the iPhone build could proceed.

What problems were found during the work?

Visual review caught icons that obscured labels in dark mode and flaws in automated film recolouring. The publisher also reported eight web-app test failures that were not caused by the redesign, plus five Android findings scheduled for version 0.5.1.

Were all 228 film runs successful?

No. The publisher says production was paused after round nine, when 208 of 228 runs were complete and 38% of jobs were failing. It says the 76 failed jobs were not billed; the supplied account does not give a full breakdown of all costs or final film status.

Why does the publisher call it a two-day relaunch?

The execution took place on Oct. 2 and 3, 2026, from an approved plan to the live website. The publisher notes that an interim design draft dated Sept. 26 already existed, so the design work did not begin from scratch on Oct. 2.

Source: ThorstenMeyerAI.com