🔍 Read the full analysis: Can AI Agents Transform A Brand’s Look On A Tight Schedule? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Gewerkton says AI coding agents implemented its HORIZON redesign across its apps, website, icons and promotional films on Oct. 2–3, 2026. The company’s founder approved plans and releases, while human visual checks caught problems automated tests missed; the source does not provide a complete project cost or independently verified results.
Gewerkton, a construction-site documentation tool built in Iffeldorf, Bavaria, says it replaced the look of its web and phone apps, Studio, icons and multilingual website on Oct. 2–3, 2026, using AI coding agents to carry out implementation from written plans. Founder Thorsten Meyer approved the plans, checked results and released each part, making the project a case study in how agent-assisted production can compress a broad redesign while leaving key decisions and quality checks with a person.
The company says the redesign, named HORIZON, covered 529 website pages in 40 languages, 26 icons, the web app, the iPhone and Android app, and Studio in light and dark modes. In the web app, 281 hard-coded colour values were moved to design tokens, with 11 deliberate exceptions retained, and 47 cards were restyled. The phone app received version 0.5.0. The website also received 228 AI-generated film runs for five regions, alongside eight help videos.
Gewerkton describes the work as two days of execution, not a redesign conceived from scratch in that period: an interim visual draft dated Sept. 26 informed the effort. Agents including Codex and Claude worked in a separate copy of the code on their own branch. A written plan listed decisions, alternatives and recommendations; Meyer approved its recommendations, reflecting questions raised in discussions about interfaces for AI agents. Work was not merged or released until he had reviewed and approved each part.
The reported checks combined automated tests and human inspection. The phone app passed 853 of 853 tests and Studio 2,158 of 2,158. The website passed seven test suites and was checked in five regions at three screen widths. The web app was visually examined across 200 views in four languages, at three widths and in light and dark modes. Its test suite still had eight failures in seven files; Gewerkton says six failures also occur on the main branch and one is order-dependent, so the redesign did not cause them, but did not resolve them either.
Where Human Review Made the Difference
The project illustrates both the speed and limits of using coding agents for a coordinated brand update. Agents can make broad changes across a codebase and prepare many localized assets, but Gewerkton’s account says the workflow depended on human approval gates, visual review and decisions about what to do when output fell short. It is a company-reported example, not independent evidence that the same schedule or results will hold for other products.
Several problems emerged only when people looked at the output. The first icons appeared on white tiles that obscured nearby labels in dark mode; automated tests had not flagged the issue. A style rule also displaced the icons from their intended column. The company says it corrected the icons and layout and released the fix that evening. In the Android test on a Pixel 10a, larger type and padding caused camera-screen buttons to break mid-word. Gewerkton says five major findings came from that test, none considered blockers, and all were scheduled for version 0.5.1.
The experience matters to teams weighing rapid redesigns because passing tests did not mean every visual detail was right. Here, human inspection caught interface and media defects that automated checks missed, while explicit approvals kept an agent from independently publishing sensitive or paid work. That division of responsibility—not just the agent count or turnaround—is central to understanding the reported result.
Products Worth Considering
Amazon Kindle Paperwhite (12th Generation, 2024), 16 GB, 7-inch Display, Jade
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
The Plan Behind the Two-Day Run
Gewerkton says the HORIZON rollout began with a plan that identified open choices, gave options and made recommendations. Meyer accepted the plan as recommended, then reviewed the implementation step by step. The agents’ isolated branch kept their changes separate from the main code until release. Some actions remained strictly human-controlled: films showing the apps were held until Meyer approved the updated interface on Oct. 3, and paid film reruns required his explicit permission.
One release decision concerned accessibility. The original cobalt colour measured 4.38:1 contrast against the light background, below Gewerkton’s stated working threshold of 4.5:1 under WCAG 2.1 AA. The team darkened it to 4.80:1 rather than shipping the planned shade. Apple also blocked the iPhone build until its latest Program License Agreement was accepted, an action only Meyer could complete.
The film work was less straightforward than the headline schedule might suggest. Automated masking was used to remove red, which was no longer part of the brand palette, from helmets, jackets and emblems. It missed red in shadows and warm backlighting; visual review found partly recoloured objects, and 24 files were cleaned. Ten regional versions with remaining red were temporarily switched to the European cut; the source says nine were later reshot with approval and checked. One wider mask turned a person’s lips blue, while other attempts affected a face or sticker, and those outputs were discarded.
“The two days cover the execution, from the approved plan to the live website.”
— Gewerkton’s project account
What the Project Account Leaves Open
The details come from a disclosure by Gewerkton’s publisher, which says it builds the product and documents what it learns. The account provides internal test counts and production records, but no independent audit of the redesign, its quality or the claimed schedule. It also does not give a complete accounting of staff time, agent costs or the total project budget. The source mentions an actual film-production figure of about $134, but the supplied material ends before explaining its scope, so it cannot be treated as the total cost of the redesign.
The account says production was paused after round nine, when 208 of 228 film runs were complete and 38% of jobs were failing. It reports that 76 failed jobs were not billed, but does not specify the denominator or a full breakdown of the failures. Nor does it explain how the remaining work was scheduled against the two-day window. These details limit comparisons with other AI-assisted redesigns.
Some issues were still pending when the account was written: the five Android findings were assigned to version 0.5.1, and the web app’s existing test failures remained. The supplied material does not give a later release date or confirm that those items have since been resolved.
Pending App Fixes and Film Checks
Gewerkton’s stated next software milestone is version 0.5.1, which is scheduled to address the five findings from the Pixel 10a Android test. The account does not provide a publication date. The web app’s eight failing tests also remain part of the reported status, although the company says most fail on the main branch independently of this redesign.
For the media, Gewerkton says nine regional film runs were reshot after the first visual review and passed another check. The source describes a review process that included inspecting the middle frame of each of 157 mirrored films and watching reworked films again at 10%, 50% and 90% of their length. It does not give a final resolution for every film issue or confirm whether further reruns were required.
The account presents the rollout as a completed implementation, with human approval retained for releases, visual decisions and paid reruns. A fuller assessment of the approach will depend on whether the pending app fixes arrive, whether remaining test failures are addressed, and what complete cost and timing records show.
Key Questions
What did Gewerkton change in the redesign?
Gewerkton says HORIZON updated its web app, iPhone and Android app, Studio, 26 icons and 529-page website in 40 languages. The rollout also included 228 regional AI film runs and eight help videos.
Did AI agents make the redesign decisions and release it?
According to Gewerkton, agents carried out implementation from written plans, but founder Thorsten Meyer approved the plan, reviewed results and released each part. He also had to approve paid film reruns and complete an Apple agreement before the iPhone build could proceed.
Was the entire redesign created in two days?
No. Gewerkton says the execution took place on Oct. 2–3, 2026, but an interim design draft from Sept. 26 was carried into the work. The two-day description applies to implementation from the approved plan to the live website.
What problems were found during the rollout?
Visual checks found icons that obscured labels in dark mode and AI-edited films with missed red details or unwanted colour changes. A Pixel 10a test also found five non-blocking Android issues, including a button label breaking mid-word. Gewerkton says fixes for those app findings were scheduled for version 0.5.1.
How much did the project cost?
The supplied project account mentions about $134 in actual film costs, but it does not explain the figure’s scope and does not provide a full project budget. It should not be read as the total cost of the redesign.
Source: ThorstenMeyerAI.com
