What we're reading this week

A Thinking Machines engineer explains how accepting agent output uncritically wears down his feel for the systems he owns, then lists countermeasures: a personal "vibe limit", treating repeated "yes, continue" as a warning sign, typing every edit by hand, and PR views with the agent's commentary beside the diff.

An architecture for cheap models on tight rails in high-volume casework: autonomy tiers, a human-scored gold set, a confusion matrix that prices each kind of mistake, and an orchestrator that never lets models write to systems of record. Yegge contrasts it with his own wide-open software factory. The model-welfare sections can be skipped.

After the midterms, Thompson argues the binding constraint on AI has moved from compute, memory and power to political permission to build, so AI may stay scarce. Ten takes follow on who wins and loses, from hyperscalers and frontier labs to Apple and China, and why startups lose cheap compute.

A short account with data from a Postgres committer: with LLM bug reports, commits to the new stable branch rose 61% in a year and back-patches 74%. Fixing them used up review time, which slowed decisions on which late features to revert and delayed the release. Finding bugs got cheap, but fixing them still runs at human review speed.

Rothman on "mattering" as AI, aging and being replaceable make people feel redundant. Drawing on Junger, Sargeant's The Dignity of Dependence and MacIntyre, he argues that being needed comes from ties to other people and being ready to help, not from output nobody else could produce.

The hosts break down the latest SlopCodeBench results on agent code quality: strict pass rates as requirements change, medium effort beating x-high for Opus 5.5, and style rules turned into skills with compiler back pressure. The argument is that maintainability compounds and agent speed and token cost don't. A video; the newsletter recap carries the substance.

Lawson tests claims that some frameworks are "better for agents" by building the same app five times in each of 12 frameworks with Opus 5.5 and measuring cost, tokens, lines of code and Lighthouse scores. Hono with htmx and Astro came out cheapest and leanest, Angular costliest, and agents often ignored the frameworks' official MCPs and skills.

A profile of Terence Tao moving from AI optimism to alarm as AI solved Erdős problems and Navier-Stokes was claimed. It covers the flood of proofs, collapsing peer review, "proof digestion", and how to train graduate students when their formative problems get solved in hours.

A Salesforce AI research engineer compares his AI use with his account from a year earlier. He now runs five or six agents and spends about $3K a month on tokens, but his best work was a few hundred hand-written lines that replaced 4,000 lines of vibe-code. He also reports that AI misled his novel research and that juggling agents is tiring, and he includes a personal AI policy.

Larson on Imprint's verification loop (a required verifications file for every PR, a daily maintenance pass and a feedback tool) and on whether to centralize a team's AI tooling or let people explore. He admits doubts about centralizing too far.

The winner of an essay contest compares his own PR data from Bloomberg, PostHog and the AI-native startup Tano: 11x more PRs, no human code review, and migrations in weeks instead of a year. He is candid about the costs (thrashing, lost shared context, a bus factor of one, exhaustion) and the guardrails he added.

Recent work of ours. In one of our meetings, Claude turned the live transcript into a shared board that the room commented on while they were still talking, and the comments grew it a diagrams page, a deck and an animated 3D page. We write up what it took and where it fell short.

Fisher uses the Charge of the Light Brigade, the tapper experiment and Toyota's testimony to show how intent decays as it is passed along, then recommends Bungay's backbrief and Drucker's manager's letter as checks on what actually arrived.

Patrick Morgan of Sublime Security looks back on an agent-driven prototyping codebase his team adopted on its own: an AGENTS.md with layered context, per-contributor folders with auto-approved PRs, personas the agent uses to critique designs, and prototype components kept as copies of production that the agent keeps in sync. He says agents have made that duplication cheap to maintain.

Beck explains why agentic work can't be fired and forgotten: software has no fixed done state or fixed implementation, agents report success that isn't real, and they pile up complexity they never reduce, so the human check-in has real work to do.

A shared Git repo with AGENTS.md, skills and hooks, rolled out to 60 to 70 PMs, each with a private sub-repo. Skills raise the floor on work that usually gets skipped, and bug bashes in the browser follow explicit test plans. The write-up is candid about the gaps: no benchmark, no telemetry, instructions pruned as models improve, and a human gate kept on automated experiments.