Teams building with AI agents are turning each failure or correction into a permanent check, rule or shared record, rather than fixing one result at a time. Six accounts, from engineering leaders, designers and a design team at Cisco, describe that shift from different jobs.
Cate Huston (4 Nov · 01), an engineering leader who writes the blog Accidentally in Code, found a menu button floating in the wrong place on a phone screen. Her team had a design system and automated screenshot tests, but the tests only captured desktop width. She fixed the bug, then turned on phone-width screenshots, which found nine more layout bugs. She added a check that fails the build when elements overflow. She also backed a rule in the AI agent's instruction file with automatic checks, because, in her words, "An instruction is a suggestion."
Neil Kakkar (4 Nov · 02), a founding engineer at Tano, an influencer marketing agency that builds AI agents, won The Pragmatic Engineer's essay prize with data from three jobs. He now merges more than 200 pull requests a month, about eleven times his rate at the analytics company PostHog. At Tano an AI review agent is usually the only review a change gets, and each new way production breaks is added to that agent's instructions. He reports the costs plainly: a currency bug, a payment made twice, constant task switching and the loss of shared knowledge that reading colleagues' code used to give.
Jack Franklin (4 Nov · 03), a senior software engineer at Google writing in LeadDev, says one engineer with an AI model can now plan work across the whole technical stack. His job as tech lead has moved from coordinating specialists in advance to keeping the team's work consistent. He describes setting rules and tools for the agents, and making sure a domain expert reviews any AI-guided plan outside an engineer's own area. He also insists that AI-generated mock-ups do not replace a designer.
Stephen Patterson (4 Nov · 04), a product designer at Deck, a company that builds software agents which operate websites the way a person would, watched thousands of agent runs. The agents stalled where people also struggle: two buttons both labelled "Continue", menus that appear only on hover, sessions that expire and payments that never confirm. People absorb these problems by guessing; agents cannot. He now asks five questions in every design review, such as whether someone who cannot guess could finish the task.
Alejandro Marrufo and Casey Tong (4 Nov · 05), a UX designer and the head of UX at Cisco DevNet, Cisco's developer programme, described rebuilding an internal publishing tool on a tight deadline. An AI coding agent built a working interface from user stories and backend APIs, but it did not follow the team's visual language. Redrawing it in Figma lost interactions and left two versions with no clear current one, so the designers refined the working HTML directly with an AI coding tool. Each version now goes to engineers with its design system and a changelog, and a coding agent applies the changes to the real project.
Jakub Otcenas (4 Nov · 06), a designer at the design studio Diorama, wrote in Codrops about building the studio's design system with AI agents. He reviews an agent's work by leaving comments on the design canvas, explaining why each choice matters as he would to a junior designer. At the end of each session the agent turns those comments into written rules and skills, kept in a file called CONVENTIONS.md, so the next session starts from them.
In each account, the useful human work moved from checking one output to changing what produces or receives the next one. Huston's build checks, Kakkar's growing list for the review agent, Franklin's rules and expert reviews and Patterson's review questions all make a failure harder to repeat. Cisco's changelog and Otcenas's rule book do the same for design corrections, so a designer's fix reaches the next version without being explained again. The inference, which is not any single author's claim, is that teams should count which failures became checks, not only how many changes shipped.
Kakkar's own numbers show that a small team can drop human review and still ship quickly, while fixing bugs after release. For many teams, building and tuning these checks may cost more than the occasional bug they prevent, as Huston's two weeks of flaky screenshot tests suggest. The Cisco and Diorama accounts each describe one internal tool or one small studio, with no measured results.
Engineering leader Cate Huston traced one misplaced mobile button to tests that checked only desktop width, and argues that engineers now fix the system that let it through.
Founding engineer Neil Kakkar shows that at Tano an AI review agent replaced human code review, raising his output elevenfold while shared knowledge and focus suffered.
Jack Franklin, a senior software engineer at Google, says AI lets one engineer plan across the stack, so his tech lead job moved to keeping work consistent and reviewing plans.
Product designer Stephen Patterson found that AI agents operating websites fail on the same poor design that people quietly work around, and now uses those failures in design review.
Cisco DevNet's UX team stopped redrawing an AI-built interface in Figma and made the working HTML prototype, with a changelog, the handoff to engineers.
Designer Jakub Otcenas of the studio Diorama builds its design system with AI agents, turning his comments from each session into a written rule book the agents follow.