Escaping the No-Code Platform by Building a Software Factory

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.

Cate Huston once ran about half of DuckDuckGo's engineering organisation and now leads engineering at Twill, a hiring marketplace where members refer people to jobs. Twill had outgrown a no-code platform, and Huston and one other engineer rebuilt it with AI coding agents, launching on 2 August after four months. She describes the system they built, which she calls a software factory: agents write most of the code, a merge queue lands it, and automated checks decide what gets through.

• After each bug, the team names the class of error, adds a check that detects it, then changes the code so the error cannot be written. A bug that mixed dollars and cents led to one money module and a type that rejects raw dollar amounts.
• Every agent must break its own fix, show a named test failing, restore the fix and put that evidence in the pull request.
• A hook stops the AI agent merging any database migration, and a person takes responsibility for migrations and releases.
• In September the team merged 779 pull requests with one emergency release, and its spending on automated builds and tests was higher than its spending on AI.

Huston suggests that existing teams start by turning past incident reviews into automated checks, and by tracking how long pull requests stay open and what builds cost. Her figures come from a team of two on a new codebase, and she says retrofitting existing teams is much harder.

Derived by Working Surface from the article; more in the Playbook. Source line: AI will make the same mistake endlessly.

28 October 2026: as AI agents make code look finished sooner, teams are moving the human check to before the build and keeping a person accountable for the changes that are hardest to undo.