What we're reading this week

Dan Luu checks the claims against real repos and benchmarks and argues against "meat proxy" work, where a person runs an LLM in a loop without judgment. He follows it to the end: if oversight with your brain off is enough, your employer has no reason to keep you.

Breck used AI heavily on an academic paper without letting it write a line, and explains why AI-written documents fail the reader. The person who prompted the model has the context that made the output make sense. The reader gets the conclusions without that path and has to read every line to rebuild it.

Maggie Appleton, a design engineer at GitHub Next, on sketching in a paper notebook instead of describing ideas to Claude Code, the fatigue of planning with agents all day, and writing verification specs instead of reading the code. She also argues for new artifacts where people and agents can work from the same page.

Interviews with the people rebuilding Windows around agents as ordinary OS users: agent identity through Entra ID, an on-device registry of MCP servers, isolation for agent tools, and running models locally.

Pointing coding agents at nanosecond-level optimization. Keep the design loop and the implementation loop separate, since agents are good at only one of them. Do a delete-only pass, then rebuild in a fresh context so the model can't fall back on old patterns it can still see.

Husain and Shankar, drawing on work with more than 50 AI companies, on why teams skip finding their errors and end up measuring the wrong thing. The idea to keep is criteria drift: you find out what good means by reviewing real outputs, so the eval changes as you use it. It comes with a free evals plugin for coding agents.

Derives a measure of coupling and cohesion from first principles and shows that SOLID, DRY, the Law of Demeter and KISS all restate the same two variables. The author is explicit about where the model stops working.

Thompson uses Christensen's integration and modularity theory to read Anthropic's case for pacing the frontier. The safety framing also happens to answer the labs' problems with capability, product, pricing and capital as models become good enough. It continues last week's pacing debate.

Agents now write most of Linear's tests, and CI became the constraint. The post gives the numbers: runner-minutes, shard configurations, caching, and timings before and after the rework.

Why AI is causing a CPU shortage: reinforcement learning and agents running tools and tests are CPU-heavy, spot pricing has disappeared, and chip capacity is shifting toward high-bandwidth memory. With input from turbopuffer's CEO, Claude Platform's Katelyn Lesse and data from Uber. Capacity planning for CPUs is back.

A design for a randomized trial that would show whether AI speeds up delivery, with delivery speed kept separate from productivity and four measurement traps named. Shore cites a METR study where engineers overestimated AI's speed benefit by about 50%.

Ptacek uses the LLM as a copyeditor and never as a ghostwriter, under two rules: never keep its words verbatim, and ignore its praise. His reason is the reader. Read it beside Breck's piece in the first section.

A case study inside a 1,200-person Danish software company a year into its AI rollout, from 21 interviews and 12 follow-ups. It names a "verification tax" (engineers still answer for code they didn't write, so they inspect all of it) and finds the emotional cost tracks how much of a role's craft the AI took over. The quote is one engineer, asked how he handles the added mental load.

A talk on the long tail of software that big vendors won't build because it doesn't scale, and on "barefoot developers": people outside the profession making software for an audience of one, named after China's barefoot doctors. Grounded in working prototypes.

Elena Verna's notes from more than a year at Lovable: a flat structure, individual contributors carrying high-impact work, fewer meetings, and management rethought. The pull quote is the sharpest line on what management was for.

A walkthrough of Warp's agent pipeline from Slack request to pull request, with the mechanics shown: a "human interactions per PR" metric, an LLM judge scoring every run, and an observer agent that rewrites the factory's own agent definitions after 20 to 25 recurring failures.

A research-based framework for the quality of the context you give agents, scored on clarity, actionability, fidelity, efficiency and security, with a real incident behind each. The best example is in the quote: the same instruction gets better by moving it.

Anthropic's field guidance on the work before agents touch a migration: define the target, write the "certificate" of checkable evidence each change must carry, set a tiered promotion policy, stage the prerequisites, then run on one partition before scaling. It's vendor writing, and it's practical rather than a sales pitch.

Will Larson's account of Imprint's first software factory, rough edges included. A chain of Linear, Notion, Datadog and Snowflake decides what the agent may work on next, and building it showed him how much state he had been keeping in his own head.

A GitHub Next talk: one developer running a fleet of agents gets faster, and the team doesn't, because agreeing on what to build is the bottleneck. PRs and issues are the wrong tools for that, since review comes after the decision. She demos a collaborative multi-agent tool built on the idea.

From a two-person startup building with agents: merge queues, CI that takes 13 minutes in the monorepo and 2 in an extracted service, and an agent that delegates work across repos. Once agents review code quickly, the repo boundary becomes the thing you review.

Fisher ties together soccer analytics, the Halo Effect and Army after-action reviews: judging by outcomes erases the difference between good process and good luck. The fix is a habit. Write down the hypothesis, and what would prove it wrong, before the result comes in.

Fisher again, with NUMMI, Pronovost's ICU checklists, Project Aristotle and research on analysts who changed firms. Much of what looks like individual capability belongs to the system the person works in, and performance reviews that count one person at a time miss it.

How Stripe spreads expertise without centralizing it: APIs that work from any surface, an Agent Studio where domain experts own and tune their own agents and skills, and a shared environment to run them in.