What we're reading this week

Vaibhav no longer reads most of the code he ships, so watching the system run becomes the only way to understand it; a case for typed, queryable traces and for reading rising errors as users pushing against a boundary that used to be invisible.

A 5,920-person survey where AI-identity ("amplified" vs. "diminished") now predicts career sentiment more than role or tenure; the real fear is the squeeze, not replacement, and the apprenticeship rungs are vanishing.

Looks like a tool roundup, argues the opposite: the tool doesn't matter, judgment and learning do, with concrete agentic patterns (define "done" as pass/fail checks up front, a separate grading agent, config-as-scars).

Understand agent-written code not to verify (agents self-verify) but to participate; the concepts in your head are what let you drive the next loop. Introduces "cognitive debt" and quizzes as a speed regulator.

Claire Vo builds a repeatable benchmark harness with Claude Code and finds her taste-based rankings land almost exactly opposite the LLM-judge leaderboard: automated judgment as an easy grader that misses what the human eye catches.

Scott Alexander runs AI forecasters on live questions at ~$8/query and maps where human judgment still holds versus where it's now commoditized, including catching himself reject a forecast for challenging his worldview.

From our own desk: Ben pushes past the "more planning, more review" consensus to the half it leaves out. The bounded task now belongs to the agent, so each builder's appetite has to grow to whole features, and the controls (self-testing agents, real preview environments, fresh-context review) have to grow with it until the role reads builder, not developer.

Built from 50+ interviews: once AI makes resumes and applications near-free, the inbound channel collapses into noise and trust, network, and taste become the scarce moat.

Field notes from a Fowler-hosted Thoughtworks retreat on where value migrates when code generation is cheap: quality becomes automatable, high-level design and specs stay human, "software engineering becomes specification engineering."

A practitioner documenting a real build, warts and all: an "agentic user in the loop" pattern where Claude Code iterates with a persona agent over Slack and ships harness features overnight, plus a concrete credential-compartmentalization architecture for the "lethal trifecta."

The "agent prompter to agent manager" shift: a cloud Symphony+Linear software factory with per-task token accounting as a health metric, and a Pokémon-arbitrage demo of AI collapsing the information advantage scale used to provide.

Thorsten Ball defines ownership as owning a problem end-to-end, with a mental checklist that both encodes small-team leverage ("no PMs, no QA... we're small but great") and transmits capability to juniors without flattening them.

Exactly the cross-domain move the reader prizes: oral epic bards composed like jazz musicians (Homeric formulas as improvised "licks"), grounded in Parry & Lord's fieldwork: what tacit, living craft is lost when a sung tradition is formalized into text or film.