Teams that let AI agents write most of their code are finding that checking that code is now the slow and costly step. Six pieces published on 6 October show people deciding where human judgement must go: in code review, in design systems and in what gets released.
Emma Bostian (7 Oct · 01), an engineering manager at Spotify, the music streaming company, argues in LeadDev that AI coding tools have not made delivery faster. In many teams, pull requests now arrive faster than engineers can understand, test and approve them. She describes non-technical colleagues who open pull requests with AI and then pass each review comment back to an AI agent without understanding the problem. Citing Harvard Business Review research, she notes that each piece of low-quality AI work takes nearly two hours to handle. Her remedy is that the person who starts a change must be able to explain it, test it and review it before asking a colleague to do so.
Cate Huston (7 Oct · 02) is the part-time chief technology officer of Twill, a company that connects employers with senior candidates through referrals. Twill has two engineers, and AI agents write most of its code. Her scripts show pull requests rising 24-fold from April to September, while automated safety checks grew from 12 to 36 percent of them. Same-day emergency fixes fell from about ten in August to one in September, but the median pull request stayed open 3.5 hours instead of 1.2. Huston accepts that trade because incidents are very expensive for a team of two engineers.
Katie Fraser, Agnia Sergeyuk and Ilya Zakharov (7 Oct · 03) are researchers at JetBrains, a maker of programming tools. With researchers at Lund University, they asked 17 practitioners in four workshops how a tool for reviewing agent-written code should work, then surveyed 43 software professionals. They point out that an AI model presents every line with the same apparent confidence, and a reviewer cannot ask it why it wrote something. Their proposed tool shows an overview first, then ranks files by risk, and only then opens individual pieces of code. The work is a design proposal from a tool vendor, not a tested product.
David Heinemeier Hansson (7 Oct · 04) is co-owner and chief technology officer of 37signals, which makes the project tool Basecamp, and the creator of Ruby on Rails. Two weeks ago he told the Rails World conference that programmers will no longer write most code by hand. When he asked who still wrote much code by hand each week, only a handful raised their hands. He describes AI agents as fast coworkers whose choices a programmer may sometimes disagree with. In his view, refusing to work with them is not a viable career path.
Jenny Xie (7 Oct · 05), an editor at Figma, the design tool company, reports how three customer teams use the AI agent inside Figma. At Uber, Staff Product Designer Ian Guisard built agent skills that diagram a component and map its colours to design tokens. Documentation that took several people months is now published by one designer in an afternoon. At Granola, which makes a note-taking app, a designer has the agent place meeting feedback on the design canvas. The post is Figma's account of its own product.
Petras Baukys (7 Oct · 06) writes for Muzli, a browser extension that shows designers new work. He argues that long research phases made sense when building the wrong thing cost months, and that AI has removed most of that cost. Muzli now builds small changes on the day an idea appears and releases them to half of new users. The team writes down the sample size and the reading date before each test starts. Baukys notes that a wrong product still costs users and their trust.
All six treat human judgement as the scarce resource once AI agents do the making. Bostian makes authors answer for their own changes, Huston adds automated checks, and the JetBrains researchers design tools that point a reviewer at the risky parts. Hansson, the Figma customers and Baukys describe the same shift outside code review: people choose and check what AI agents produce. The inference, which is not any author's claim, is that a team gains from AI agents only as far as it plans how that attention is spent.
Huston's numbers come from a company with two engineers, Bostian's essay does not describe her own team, and the JetBrains framework has not been tested in a shipping tool. Hansson offers an opinion without figures, and the Figma and Muzli pieces each promote the publisher's own product. Larger teams may find that review was never their main constraint.
Emma Bostian, an engineering manager at Spotify, argues that AI coding tools moved the delivery bottleneck from writing code to reviewing it, and that authors must own their changes.
Cate Huston, part-time technology chief at Twill, measured six months of agent-written pull requests and found that automated safety checks cut emergency fixes but slowed merges.
JetBrains researchers propose that tools for reviewing agent-written code should show an overview, rank files by risk and only then open code, based on workshops with 17 practitioners.
David Heinemeier Hansson, who created Ruby on Rails, argues that AI coding agents now write most code and that programmers who refuse to work with them risk their careers.
Designers at Uber, Granola and Atlassian describe using Figma's AI agent to document components, collect review feedback and add motion, in a post Figma published for the agent's general release.
Petras Baukys of Muzli argues that AI made a wrong prototype cheap, so design teams should build first, release to half of users and measure before deciding.