AI agents can now do much of the work of building software, and teams are deciding on purpose which parts people keep. Four accounts published on 29 October show engineers and leaders keeping judgement, review and measurement with people while agents do more of the rest.
Rachel Laycock, chief technology officer of the software consultancy Thoughtworks (30 Oct · 01), asks why many of the company's best developers hesitate to let AI agents do all the coding. Giles Alexander, its regional chief technology officer for Europe, ran a week in which senior engineers, infrastructure consultants and product people built an airline application without writing any code themselves. The teams spent about two thirds of their time at the whiteboard and one third on the tests, contracts and delivery pipeline the agents worked against. Even so, the build was often broken. Laycock argues that the caution comes from knowing that built software is not yet software that works in production.
Sean Goedecke, a software engineer who works on GitHub Copilot, GitHub's AI coding assistant (30 Oct · 02), lists the routine tasks he now hands to AI agents. They start his development server, test the main user paths, watch metrics and logs for hours after each deploy, and gather second opinions from other AI models. A scheduled agent reads his overnight messages each morning and researches the problems in them, because his American colleagues start work hours before him in Australia. He still writes his own messages, makes his own design decisions and reviews code himself, because he is responsible for that work.
James Stanier, chief technology officer of the software company Nordhealth (30 Oct · 03), writes in LeadDev that cheap AI usage for engineers may not last. His team is on a flat-rate plan, but he estimates the same use would cost about ten times as much at per-token prices. He asks every engineer to know how to read their own usage and which AI model suits which task. Managers see their own team's figures, and everyone else sees department figures without names. He warns against ranking engineers by usage, because people then use AI for its own sake.
Microsoft Digital, Microsoft's internal IT organisation (30 Oct · 04), describes in a playbook written by Jason Kellington how it moved from scattered AI experiments to six connected employee councils. The early experiments had produced duplicated work, uneven governance and gaps in reporting. The councils now decide which AI projects matter and ask teams to fix a process before automating it. They also review each project for responsible use before work starts and require a measured baseline before launch. This is Microsoft's account of its own practice, and Microsoft sells the AI tools involved.
In each account, people decide in advance what AI agents may do and how their work will be checked, whether by tests, by a person or by a measured baseline. The inference here, not any one author's claim, is that these checks let a team hand more work to agents. They keep the team able to tell whether that work is right, affordable and useful.
Three of the four accounts describe one person's or one company's experience, and none reports results measured in production over time. The Thoughtworks trial often ran with a broken build, which suggests the checks may not yet keep pace with the agents.
Thoughtworks' chief technology officer Rachel Laycock says experienced developers distrust AI agents because built software is not yet working software, and they should help design the new practices.
Sean Goedecke, a GitHub engineer, hands AI agents his routine engineering chores but keeps his own messages, design decisions and code review, because he answers for them.
James Stanier, chief technology officer of Nordhealth, tells engineering leaders to make AI usage visible and match AI models to tasks before per-token billing raises costs about tenfold.
Microsoft's internal IT organisation replaced scattered AI experiments with six employee councils that choose projects, fix processes before automating them, review responsible use and measure results.