AI agents now let designers and engineers produce changes quickly, and the work left for people moves to review and responsibility. Five pieces look at who reviews agent-built work, who answers for it, and what an interface should promise.
Leonardo De La Rocha (21 Sep · 01) leads designers at a company that makes software for clinicians, and writes the newsletter Design Outcomes. In one week, one of his design leaders and one of his design directors each used an AI coding agent to submit a code change. The trouble came in the review, where the design leader could not understand the comments on his own change and engineers did not want to review a designer's work. De La Rocha's position is that the tech lead or engineering manager of the team that owns the code reviews each change, to the usual standard. In his words: "It doesn’t matter who did the thing. It matters who owns it, who reviews it, and who merges it to production."
Pavel Samsonov (21 Sep · 02), who writes the product newsletter The Product Picnic, argues that people now treat the AI model Claude as a colleague. Someone asked about a document they handed over can answer "I don't know, Claude wrote it," and nobody answers for its contents. Errors then pass through several hand-offs before anyone checks them. Samsonov holds that a model has no agency, so the person who hands the work over stays responsible for it.
GPS (21 Sep · 03), a developer experience advocate at GitHub, recapped a GitHub Podcast episode that asked whether people still need to read code written by AI. Her answer is yes, because developers remain responsible for it, but the effort should follow the risk. A rewrite of a live product's sign-in code deserves a different review from a styling experiment. Her rule is to "review until you can explain and own the outcome."
De La Rocha (21 Sep · 04) also wrote up a silent critique, a review in which everyone writes feedback before anyone speaks. A designer adding an AI button that explains a chart had drafted loading text such as "looking at your data". He pointed out himself that the text described several steps, while the feature makes one call that returns in about five seconds. De La Rocha concluded that the loading state should promise little, such as a plain spinner, and a reviewer can make this check without reading code.
In a third piece, De La Rocha (21 Sep · 05) describes two AI chats in his product: a vendor's support bot and an assistant built by the product teams. Clinicians typed questions about features into the support bot, which could not answer them. He changed his mind twice in one week and settled on one chat that gains skills as each account grows. He concludes that his earlier positions kept two chats only because two teams owned them.
Agents make changes cheap to produce, and each piece puts the remaining work on people: reviewing the change, answering for it, and deciding what it tells the customer. De La Rocha and GitHub agree that the review belongs to someone who can explain the change and answer for it.
Three of the five pieces are one design leader describing a few weeks of practice at one company. Making the tech lead the default reviewer may also move the delay onto the busiest person on the team without removing it.
When designers at Leonardo De La Rocha's company began making code changes with AI agents, the review broke down, so he set out who reviews them and to what standard.
Pavel Samsonov argues that treating an AI model as a colleague lets people hand over work nobody has checked, and that whoever hands it over stays responsible for it.
A GitHub developer advocate argues that people must still review code written by AI, with effort matched to the risk, until they can explain and own the result.
A designer noticed that the loading text of his AI feature described several steps while the system made one quick call, and his design lead argued for a simpler indicator.
Leonardo De La Rocha changed his mind twice in a week about how many AI chats his product should offer, and chose one because customers wanted one place to type.