What we're reading this week

Fournier pushes back on the idea that AI ends management. Her worry is what it does to the managers who stay: dashboards and summaries give them visibility in place of knowing how the system works. Read it against Mollick's The Dot and the Swarm further down, which argues the other side.

A study that interviewed the same developers in October 2024 and April 2025. Delegating to AI moved their attention toward checking and supervising, and they reported the same productivity while their experience of the work got worse. The study predates current agent tooling.

The headline is about coding by hand dying, and the piece is better than that. Orosz covers 37signals moving to Rust, DHH's argument that abstractions were the price we paid for repetition, and concrete cases of quality slipping.

The Cockroach Labs CTO and GIMP co-creator on how AI multiplies his own work. Two numbers stand out: non-engineers at the company built about 1,000 internal apps in a few months, and a former Google readability reviewer predicts we will stop reading code.

Beck splits every field into visible work (features, proofs) and invisible work (understanding, simplifying, finding the abstraction). Machines take the visible part, so people's value moves to the part nobody credits.

A guest post on Terence Tao's blog by the cryptographer Amit Sahai. When AI makes mathematical work cheap, society needs more mathematicians, not fewer, because keeping people in charge takes communities that understand the reasoning.

Training pushes models toward the well-supported middle of a field and away from the insight that breaks with it. So what AI brings is stamina, and the ideas that move a field still come from people.

A formal-methods practitioner answers the hope that formal verification will make agent output trustworthy. The hard part is deciding how to formalize what you care about, and many of the properties that matter can't be expressed at all.

A concrete habit: ask a confirming question about every 30 seconds, until you can build the system in your head. Goedecke applies the same habit to questioning agents.

A rebuttal to "coding is solved". Accountability needs understanding, the code stays the source of truth, and how much of it you must understand depends on how much risk you can carry. The tone is snarky; the argument holds.

Rebuilds the Academic sceptics' method, from Cicero and Carneades, of matching scrutiny to the stakes. It's a model for calibrated confidence, strong on the reasoning and thinner where it applies it to science and democracy.

A second look at last week's enzyme result. A biologist worries the model was led there by his team's unpublished work rather than reasoning its way to it, which puts the discovery-or-retrieval question plainly.

Britton Russell from our team on what happens when a team's output outruns its capacity to review it: the most experienced people stop building and clear a queue that refills overnight. He names five signals, proposes one measure (senior attention consumed per unit of net value), and puts the fix in the environment before the people: context the tools read on every run, automated checks on every change, and human review for what the checks stop.

A survey of 308 engineering leaders at LDX3 New York. The gains from AI aren't reaching individuals, speed has costs further down the line, and many leaders aren't sure what their job is any more.

CodeScene refactored 300,000 lines of Street Fighter 3's C code in days for about $4,000 with Claude Code and Opus. Frame-hash tests decided whether the game still behaved the same, and the interview covers what people still had to do.

Ganesh has built a forward-deployed engineering function three times in a decade. Now that drafting a company's operating model costs almost nothing, he argues the valuable part is the understanding you get from being corrected on site.

A retrospective, mistakes included, on building an AI product's quality bar with Claude Code. Engineering capacity used to limit what got built. With that limit gone, human judgment written down as evals has to take its place.

Mollick reverses his earlier view that agent swarms would need carefully designed human management. Much of management solves problems people have and agents don't, and when organizing gets cheap, the list of things worth attempting grows.

Aggregation Theory applied to agents. When doing things is abundant, deciding what to do becomes scarce, and power goes to whoever owns the agent people ask, such as Meta's Muse or Microsoft's Copilot.

Why tokens keep getting cheaper (GPU efficiency, mixture-of-experts models, the Jevons paradox, oligopolies held together by switching costs), and what becomes scarce once they are.

College bundles promises that don't fit together: a credential, a coming of age, teaching and research. AI and funding cuts are pulling them apart, starting with what a credential is worth when coursework can be done for free.

Thompson uses data to show effortful, hard-to-coordinate social time losing to frictionless time alone, and argues hosting is becoming a luxury good.

Ten ways a strength becomes a trap, and the "saveable moment": a tripwire you set in advance so you notice while there's still time to change course. It complicates the usual advice to lean into your strengths.

An old speech that holds up. Gardner argues against settling into fixed habits and treats growth as something that keeps going, with no summit to reach.

Anthropic steered an internal Claude across more than 150 parallel Slack threads and merged over 3,000 performance changes in two weeks. The post shows the mechanics, including Valgrind benchmarks, CI ratchets that hold each gain, and the real bugs found along the way.

A guide to running agents in production: what to observe (each step's output, cost per step, tool and MCP calls, state at the point of failure), where to set boundaries, and how to put a person in the loop without losing the run. Written with Kestra, but the framework works with any tool. Paywalled.

Rasmussen and Cook's model of the accident boundary, applied to software systems, products and running a company, with two of the author's own retrospectives: the Buffer hack and over-hiring that led to layoffs.

Two Thoughtworks practitioners turn caring about colleagues into a repeatable structure: a pattern for giving feedback, a way to escalate from one incident to a pattern to the relationship itself, and feedback treated as a team's telemetry.

Arslan runs a fleet of coding agents from plan files that move through a lifecycle on disk, with coordinator agents directing the rest, and shows what goes wrong. He works at Cursor and it shows in places, but the plan-file system works with any tool.

Eleven signals for where a staff engineer's work comes from, grouped into four sources and ranked on two axes. The skill, he argues, is judging which signal to act on.