Three accounts from the last day describe the same change in how teams handle code written by AI agents. The code now arrives looking finished, so teams are moving the human check earlier and making it smaller and stricter.
Diego Quiroga (28 Oct · 01), a senior engineering leader at Microsoft, describes a team whose feature went straight from a product requirements document to code written by AI agents. The pull requests worked, but reviewers found assumptions borrowed from single codebases and missing precautions on critical paths. They asked the author for a two-page document on intent and approach, and the discussion brought the design back towards the organisation's architectural principles. Quiroga argues for short design reviews, kept for risky changes such as those that cross services or change stored data. "Code that delivers the requested behavior is not necessarily the right technical solution to the problem."
Tina Singh (28 Oct · 02), who works on a design system, describes components generated by AI that looked finished in a screenshot. In use, a card lacked a focus ring and a pressed state. Its warning icon failed contrast in dark mode, because the system had no token for status icons and the generator borrowed another one. She now settles focus, states, dismissal rules and token behaviour before asking for code, and first checks which reference the code should match. In her account, this gives the review something specific to check against, although it does not remove the review.
Cate Huston (28 Oct · 03) once ran about half of DuckDuckGo's engineering organisation and now leads engineering at Twill, a hiring marketplace. With one other engineer and AI coding agents, she moved Twill off a no-code platform in four months. After each bug, the team names the class of error and adds a check, or a code type, that stops it being written again. About a third of the team's output is now this protective work. A person must still take responsibility for every database migration and every release.
In all three accounts, the engineer's or designer's work moves from reading finished code to writing down decisions and checks before the code exists. Each team then keeps one human decision for the changes that are hardest to undo: cross-service design, design-system rules, migrations and releases.
Each account is one team's experience with no comparison, and Huston's figures come from a team of two on a new codebase. Earlier and stricter checks also cost time, and Quiroga reports that the review slowed the feature down.
Diego Quiroga, an engineering leader at Microsoft, argues that AI coding agents speed up coding but do not remove the need for a short design review on risky changes.
Tina Singh found that components generated by AI looked finished in screenshots but lacked focus, pressed and dark-mode states, so she now settles their behaviour before asking for code.
Cate Huston rebuilt the hiring marketplace Twill in four months with one engineer and AI coding agents, spending about a third of output on checks against whole classes of bugs.