🔍 Read the full analysis: A Faster Way To Relaunch? Gewerkton’s Two-Day AI Agents Project on ThorstenMeyerAI.com

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

Get hardware and tech essentials 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 the original account of the relaunch, says it replaced the look of its web app, iPhone and Android apps, Studio, icons and multilingual website on Oct. 2-3, 2026, with AI coding agents carrying out implementation from plans approved by founder Thorsten Meyer. The project matters as a concrete account of how AI agents were used in a live product release: Meyer says he reviewed and released each part, and the work also surfaced defects that automated checks did not catch.

The redesign, called HORIZON, touched several parts of the product and its public site. Gewerkton says the web app’s 281 hard-coded colour values were moved to design tokens, leaving 11 exceptions, and 47 cards were restyled. The phone app was updated to version 0.5.0 for iPhone and Android; the Studio received light and dark versions; and all 26 icons were re-rendered in Blender in cobalt and navy. The website’s 529 pages across 40 languages were also updated, alongside a new opening section using product screens.

The source account says the website upload involved 1,583 files and completed without an error. It also reports 228 AI film-generation runs for five regions, with 222 regional versions, as well as eight help videos. The account qualifies the “two days” description: an interim design draft dated Sept. 26 already existed, so the reported period covers execution from an approved plan to the live website, not the full design process.

Meyer approved a written plan that identified open decisions, options and recommendations, according to the account. Agents worked on a separate code branch; each part was held until he reviewed and released it. Films showing the product were not approved until he had reviewed the interface on Oct. 3, and paid reruns also required his explicit approval. Apple’s latest Program License Agreement had to be accepted by Meyer before the iPhone build could proceed. The source says agents could not perform that step.

Human Checks in an Agent-Led Release

The account offers a detailed, first-party example of AI coding agents being assigned a broad implementation project while a person retained release authority. It does not establish that the same approach would work for other products, or that agents independently managed the redesign: the plan, approvals, release decisions and several visual judgments remained human responsibilities.

The reported defects show why the distinction matters. An icon design that looked acceptable in one mode covered nearby labels in dark mode, and Meyer says no automated test flagged it. The team also found film recolouring errors through visual inspection, including red details missed by an automated mask. These cases indicate that passing test suites did not cover every user-facing issue in the work described.

For readers following AI-assisted software development, Gewerkton’s account provides reported measures of both scope and oversight. It says app and Studio tests passed in full, while also disclosing existing web-app test failures and problems in generated media. The evidence is the publisher’s own project record, not an independent audit, so it should be read as a company account rather than a verified benchmark of agent productivity.

From September Draft to Live Site

Gewerkton is described in the source as a construction-site documentation tool developed in Iffeldorf, Bavaria, by Thorsten Meyer, who directs AI coding agents from Codex and Claude. The publisher’s disclosure says it builds the product itself and writes about what it learns. That relationship is relevant: the project details and performance figures come from the organization responsible for the work.

The design process began before the two-day execution window. An interim draft from Sept. 26 was carried into the Oct. 2-3 implementation, and the agents worked from written plans rather than an open-ended instruction to redesign the product. When a plan did not hold up, the account says the deviation was recorded and the more conservative option chosen.

One example involved colour contrast. The original cobalt against a light background measured 4.38:1, below Gewerkton’s stated working target of 4.5:1 under WCAG 2.1 AA. The colour was darkened to a reported 4.80:1 before release. The account presents this as a project decision, not a claim that the entire product has been independently certified for accessibility.

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

— Gewerkton’s project account

Limits of the Reported Results

The figures in the account have not been independently verified in the supplied material, and it does not provide a separate assessment of whether the redesign improved usability, accessibility in practice, or business outcomes. It also does not give a full time breakdown for planning, review and correction, so “two days” should not be read as total project time.

Gewerkton says the web app’s test suite had eight failing tests across seven files. Six files reportedly fail the same way on the main branch, while one failure depends on test order and passes when run alone. The account says the redesign neither caused nor fixed those failures. It gives no broader explanation of their impact on users.

Film-generation results also required substantial review. Production paused after round nine when 38% of jobs were failing, with 208 of 228 runs completed at that point. The source says 76 failed jobs were not billed, but does not provide enough information to assess the quality or cost of the full release beyond a partial cost figure. It states that nine regional versions were later reshot and passed visual checks, while other temporary substitutions remained in place.

Version 0.5.1 and Film Corrections

Gewerkton says five major findings from an Android test on a Pixel 10a are scheduled for version 0.5.1. None was classified as a blocker in the account. One example was camera-screen buttons breaking German words across lines after the redesign increased type size and padding; the source says the word for “cancel” now wraps onto two lines.

The source also says nine film runs were later reshot with Meyer’s approval and passed visual review. It does not give a publication date for version 0.5.1 or say whether every remaining film correction has been completed. Further release details and any additional checks will be needed to establish how the identified issues were resolved.

Key Questions

What happened in Gewerkton’s two-day project?

Gewerkton says it implemented its HORIZON redesign across the web app, iPhone and Android apps, Studio, icons and 40-language website on Oct. 2-3, 2026. A design draft from Sept. 26 was already available, so the two days refer to execution rather than the entire design process.

What role did the AI agents play?

According to the publisher’s account, agents from Codex and Claude carried out implementation from written plans on a separate code branch. Thorsten Meyer approved the plan, reviewed work and released each part. Some steps, including accepting Apple’s developer agreement and approving paid film reruns, required his action.

Did all the tests pass?

No. Gewerkton reports that the phone app passed 853 of 853 tests and Studio passed 2,158 of 2,158. It also reports eight failing web-app tests across seven files, which it says were not caused by the redesign; one depended on test order, and six files reportedly failed the same way on the main branch.

What problems did the team find?

The account describes icons on white tiles covering labels in dark mode, film recolouring that missed red in shadows or warm light, and an Android camera-screen layout where larger buttons caused text to wrap. Meyer says visual inspection caught issues that automated checks did not.

When will the remaining app fixes arrive?

Gewerkton says five findings from the Pixel 10a Android test are scheduled for version 0.5.1, but the source does not provide a release date. It also does not specify whether all remaining film corrections have been completed.

Source: ThorstenMeyerAI.com