What we're reading this week

Screen-time arguments count hours. This one argues the mechanism is pace: fast media erode a child's developing capacity for sustained attention whatever the duration, backed by the author's own Children's Slow Media Project. Read it beside last week's Doomscrolling piece, which made the same argument about a whole culture.

Recaps of real personal AI setups from Refactoring's community: a terminal accountability partner, a multi-agent cloud migration, a solo embedded project. Each comes with its warts intact (fabricated citations, stale config, models quietly declining tasks), and the advice in the pull quote is the one to try.

A former architect builds a walkable 3D house over seven days and 179 prompts, and documents the loop honestly. The first result came fast. The judgment to revise it is where the week went.

Concrete agentic workflows from two designers: narrow-scoped bots, voice memos turned into Figma, bots handing work to other bots. Worth reading for how delegating production changes what a designer spends attention on.

Voss traces translating what people want into a precise spec as a separate paid role over sixty years, then argues with hiring and compensation data that it is merging back into engineering, because it is the one cost that doesn't collapse to zero.

What differentiates people once execution is cheap: deep knowledge, wide knowledge, taste, and agency. Mollick makes it a structural argument rather than career advice, and the pull quote is the whole thing in two sentences.

Seven OpenAI engineering leaders on how Codex became the backbone of internal work. Engineers drift toward product management as judgment and taste replace execution, which is Voss's argument above told from inside one company.

Tulsee Doshi on what good means when benchmarks don't capture it: precise evals written from twenty example prompts, vibes treated as real product territory, and calibrating to researchers instead of requirements docs.

Reconciles the contradictory AI-productivity studies by mechanism instead of picking a side, then argues the work moves from authoring to verification as generation gets cheap. The distinction to keep is between "it passes" and "I know why it passes".

Ronacher left a model coding unattended for 35 hours and $1,200, and shows the raw result: C edits golfed like Python, magic constants hardcoded to get past a check. His case is that what the model was trained to optimize has drifted away from code a person can maintain.

Helmer's 7 Powers re-rated for the AI era. The useful distinction is momentum, a benefit you can spend, against a moat, a barrier that lasts, worked through cornered resources, data effects and switching costs.

A retrospective on launching an AI-coded product that lists everything AI didn't solve: onboarding, positioning, the sales motion. It also names a new trap, which is that when change costs almost nothing it gets harder to ever stop polishing.

Maps the incentive network behind this week's AI-risk argument instead of taking its temperature: the Rationalist and EA funding ties, and how labs arguing for pacing regulation stand to gain a regulatory moat. Read it with the Amodei and Thompson items in Quick hits.

Majors works through her own contradictory reactions to AI-written text and comes out with a usable split: personal communication, where the value is that a specific person thought or felt something, and functional communication, where only the idea matters.

Recent work of ours. TOM Design Collective ran its interior design practice on a closed platform with no API, and handled price variances, returns and change orders in side spreadsheets. Three of us replaced it with DesignDrive in four months using agentic development, and TOM now runs live client work on a system whose codebase, data model and roadmap it owns.

Why the small fraction of people who cross role boundaries and absorb the work nobody else will do are what keeps an organization running, argued through concrete mechanisms rather than productivity advice.

Two practitioners on what breaks when agentic development scales: the bottleneck moves from code review to planning, and the hard part turns out to be the development environment, not the orchestrator.

Rand Alaboudi from our team sets out six decisions to make before choosing a model, a framework or a platform for an agentic system. They cover what the system is responsible for, where human judgment belongs, what happens when it is wrong, what has to be observable, what the infrastructure must enforce, and which choices the organization needs to keep. Each one produces an artifact you can carry into the architecture conversation, such as a responsibility boundary, a failure policy or a control map. Read it beside Nugent below, since both are about the decision that comes before the build.

Grounds "just don't build it" in real data, McGovern's Top Tasks research and a Nature study on subtraction neglect, and gives one concrete lever: show stakeholders running costs, not build costs.

As models get more capable, the bottleneck moves from how precisely you write the spec to whether you passed on your actual priorities. Give a capable model the why and it fills in the how better than you would.

Matt Pocock on the skills he uses for agentic coding ("grill-me", "wayfinder"), borrowing leading words from classic software-design books to steer agents, and memento-driven development for an agent that wakes up with no memory each session.

A measured retrospective, 160 runs per condition across 26 techniques. Naming a testing technique to an agent rarely produces its value: the agent calls the tool and skips the judgment that made the technique worth using.