AI agents now write code, design screens and summarise research faster than anyone can check by eye. Six pieces from the week to 30 October ask where that checking should sit, and most move it to the people and records closest to the real work.
Will Larson (31 Oct · 01) is an engineering leader at Imprint, a financial technology company. He found that AI agents often sent pull requests to the shared build system that failed basic checks they could have run first. His team now requires every pull request to carry a written file of each check the agent ran and each thing it could not verify. A pull request cannot leave draft without that file. Each day the team reads the files, opens issues for the gaps and updates the agents' guidance. The list of what could not be checked becomes the team's plan for better tooling.
Patrick Neeman (31 Oct · 02), a user experience designer with more than 20 years in the field, asks who should choose a design team's AI tools. In many companies, design and research operations staff, who handle licences and contracts, now make that choice. Neeman argues that only people who do the work can tell when output is wrong. His example is a researcher who spots an invented quote in a summary of her own interviews. He proposes that three to five people who use the tools every week decide which ones to adopt, and that operations handles security, contracts and budget.
Alexis Vaysse (31 Oct · 03), a product designer, has spent years keeping Figma mockups and the coded product in line. He tested designing an app with no mockups at all. He gave instructions to Claude Code, Anthropic's coding agent, and watched the real screen take shape with Astryx, a design system from Meta built for AI agents. In his workflow the designer hands over a pull request, and a developer reviews it and connects the data. He says the approach works only when automated checks reject code that does not use the design system's components.
Tanner Kohler (31 Oct · 04) of Nielsen Norman Group, a user experience research firm, examines a playbook Anthropic published in August 2026 for building software with AI agents. In it, a chain of Markdown documents passes from stage to stage, and product owners and engineers approve them. No stage or approval belongs to user experience staff, whose knowledge becomes context for the agents to read. Kohler argues that context holds only what a team already knows, and notes that the playbook never mentions user research. The advice is to claim ownership of that context, define which changes need a person's review and add signals such as task abandonment to the monitoring.
Raluca Budiu (31 Oct · 05), also of Nielsen Norman Group, argues that personal AI agents now act for people, so a journey can cross actors as well as channels. In one example, an agent recommends a flight and then asks the person to log in through its own browser, where the person's password manager is missing. The flaw shows only when the agent and the person are studied together. For Budiu, success still means that the person reached their goal.
Tim Bozarth (31 Oct · 06), a corporate vice president of engineering at Microsoft, describes software delivery as four lasting stages: plan, build, verify and observe. Agents take on more of the building and testing, and people choose the problems and judge whether the result is what was intended. A shared body of knowledge connects goals, decisions, evidence and production signals. Bozarth warns that teams changing every stage at once struggle to see what is working, and suggests improving each stage on its own.
In each piece, judgement moves to whoever can see the real output. That may be a file of what an AI agent checked, a researcher who knows her transcripts, a designer looking at the coded screen or the named owner of a team's user experience rules. None of them removes the reviewer. Each one gives the reviewer something concrete to check against. Kohler, Budiu and Bozarth add a further test: whether the result was worth building and whether people reached their goal, which a passing check cannot show.
Each account comes from one company or one person, and none reports how often errors still reached customers. An AI agent asked to report its own checks may also produce confident lists rather than real verification. Nielsen Norman Group sells courses on the subjects of its two pieces, and Bozarth's account promotes Microsoft's own developer products.
Imprint's engineering team makes every AI-written pull request carry a file listing what the AI agent checked and what it could not check, Will Larson writes.
Patrick Neeman argues that designers and researchers who use AI tools every week should decide which tools a team adopts, while operations staff handle contracts and security.
Product designer Alexis Vaysse argues that teams should stop redrawing screens in Figma and design them in code with an AI agent, handing over a pull request.
Tanner Kohler of Nielsen Norman Group finds that Anthropic's playbook for AI-built software gives user experience staff no stage or approval, and urges them to claim the context agents read.
Raluca Budiu of Nielsen Norman Group argues that design must now cover AI agents acting for people, because some flaws appear only where a task passes between the two.
Microsoft's Tim Bozarth argues that teams should rebuild software delivery with AI agents one stage at a time, defining each stage's human job and output, rather than changing everything together.