As AI agents make software cheaper to build, the scarce work is deciding what to build and recording why. A product adviser, an engineering firm, a design system team and a chief technology officer each describe moving effort from producing work to judging it and writing that judgement down.
Marty Cagan (22 Oct · 01) founded the product management consultancy Silicon Valley Product Group. He argues that AI has made building fast and cheap, yet many AI launches have frustrated customers and few have improved results. In his view the hard part is now choosing what to build, so experiments must become a core skill of product organisations. He recommends that teams own each experiment while specialists coach them, because teams that do not run an experiment take its findings less seriously. For AI products, he adds that teams should test quality with evals early and keep A/B tests for proving value.
Tenex (22 Oct · 02), a 75-person New York firm that builds AI systems for large companies, described its engineering process on the Latent Space podcast. Its cofounder Alex Lieberman and its first hire, CJ Hess, said that plans, decisions and AI agents' logs are committed to Git beside the code. An AI agent that returns to a task can then read why the team chose its approach. Engineers are paid partly by story points, and Hess estimated that output per engineer roughly doubled this year. Lieberman said the harder question is how to pay for work that is not code, such as architecture and client meetings.
Justyna Tarkowska (22 Oct · 03) is a designer who led a team's move from a Figma library to a coded design system in one GitHub repository. The team's AI agents document and audit components, and still make confident mistakes. Tarkowska's team makes each corrected agent log a lesson beside its instructions and turn it into an automated check that fails the build where possible. The team also keeps a register of deliberate exceptions to its rules, each with its reason, so agents stop reporting intended choices as bugs. Tarkowska reports that mistakes return less often after a few months, and calls that one story and not a study.
Will Larson (22 Oct · 04), chief technology officer of the financial technology company Imprint, told Gergely Orosz of The Pragmatic Engineer newsletter that the rules of engineering management have changed. In the 2010s, new managers were told never to make a technical decision again, and many solved problems by hiring. Larson now tells teams that hiring is the least interesting way to solve a problem. Orosz reports that engineering managers now often lead 12 to 14 people instead of four to eight, and that managers who still write software are in demand. Larson is hiring engineering managers at Imprint to use AI agents to rethink how software teams work.
Cagan wants product teams to test what to build, and Larson expects managers to make technical decisions again. Tenex and Tarkowska keep decisions and lessons in the repository, where AI agents read them. It is an inference, and not any author's claim, that the written reason is becoming the main thing a team owns.
Two of the four accounts come from firms that sell advice or services built on these practices. The only measures offered are story points that one firm sets for itself and one designer's impression that mistakes recur less.
Marty Cagan of Silicon Valley Product Group argues that cheap AI building makes choosing what to build the hard part, so product teams must own experiments with expert coaching.
Tenex, a firm that builds AI systems for large companies, keeps plans and decisions in Git beside the code so that AI agents can read why work was done.
Designer Justyna Tarkowska's design system team makes AI agents log each corrected mistake and turn it into a build check, and reports that the mistakes now recur less.
Will Larson, chief technology officer of Imprint, says engineering managers must now make technical decisions again and that hiring is rarely the right answer to a problem.