What we're reading this week

Maps AI-assisted engineering onto decades of supervisory-control research (Sheridan, Bainbridge): automation erodes comprehension and manual skill even while people keep noticing that something is wrong. It ends in concrete steps for redesigning the work.

A VP of Technology running parallel agents shows, with his own contribution data, how cheap implementation turns spare capacity into more work, in bursts followed by burnout.

Laycock treats developer caution as a fair question about untested ground. Her agent-only experiment ended with red builds and two-thirds of the time at the whiteboard, and she sets it beside DORA 2025 on throughput versus stability.

Adrienne Marshall from our team on what changed when we brought our CRM and knowledge base in-house. A system you own changes as fast as your team learns, your agents can reach it because you decided they should, and only your own system can hold a guardrail like a person approving each knowledge-base entry before an agent sees it.

AI-driven security risk ends the open public Internet, Thompson argues, and the response in blogs, social and software is to go small, down to apps built for yourself and a few trusted people. He reports doing this at Stratechery.

Maps AI panic onto pre-capitalist mercantilism: intelligence treated as a fixed stock to hoard, distrust of ordinary people, supply-side thinking. Then asks what demand for cognition looks like once people are freed from subsistence cognition.

Luu reports what he built with agents (a regex engine, a ripgrep fork, a Rust interpreter) and argues that old cost estimates keep us from attempting things. An appendix makes the case for being less ambitious, and he says the piece was written quickly.

How teams built AI incident investigators: a scripted workflow, then a single agent, then separate investigator and verifier agents to stop the agent agreeing with itself. Also covers the deskilling risk for on-call engineers. The free part stops at section 4.

A practitioner debate on defining an agent loop declaratively, measuring it (DORA, an automation index, online evals on recorded runs), risk-based code review, and the org chart as the bottleneck. A video; the newsletter recap carries the substance.

An Anthropic hiring manager argues that a fresh perspective plus a weekend prototype beats tenure when execution is cheap, citing two newer engineers who prototyped the Claude Code GUI in a weekend.

A twenty-year Amazon insider argues the six-page memo no longer proves clear thinking once AI can produce the document without the thinking. We could only read the opening, so the rest of the argument is unread.

Google's readability program as a giant mentoring classroom, why testing became mandatory, and the "middle group" of CS students who no longer understand code written with AI. Winters recommends grading prompts and tests instead of code.

Anthropic's research and Labs product lead on onboarding with no scaffolding, hiring for grit over titles, and replacing annual OKRs with a few stubborn bets plus one-month plans.