What we're reading this week

Mollick publicly revises his own thesis (Co-Intelligence to Co-Existence) and reflects, with skin in the game, on the cognitive cost of letting AI unstick him mid-writing, plus the genuinely fresh angle of AI as reader, critic, and gatekeeper deciding whether to recommend your work. It hits the unifying question directly.

A deep, concrete Claude Code workflow guide squarely in the AI-preferences sweet spot: CLAUDE.md as compounding infrastructure, 'let Claude write rules for itself,' skills and subagents as reusable expertise, parallel worktree sessions, three-tier Obsidian memory. It also touches how the tool reshapes the operator's mental model.

A structural analysis of why agents writing typed TypeScript or Python to call tools beats chaining bash, covering context-window output shaping, per-tool credential attachment, and the OpenAPI-spec-as-durable-investment argument, with the durable framing that tool calls are the primitive and bash vs. code mode are just implementations.

The cleanest articulation in this batch of the unifying question: agentic AI severed building software from understanding the domain, moving the binding constraint from 'can you build it' to 'can you tell whether it's right.' The domain expert has the oracle; the generalist engineer has none. Argues the durable scarce thing is a verified mental model of a real domain.

Claire Vo's warts-and-all practitioner test of Opus 4.8 documents real regressions (hallucination when stuck, struggling to orient in existing codebases, worse strategy work than 4.7) rather than benchmark hype, the kind of honest hands-on retrospective the prefs prioritize.

A book-length practitioner field guide ('what I wish someone had handed me the first time I shipped an AI feature') from a 15-year backend engineer: foundation models, prompting, eval, RAG, finetuning, agents, multi-agent, GCP deployment.

Kent Beck (creator of XP) changes his mind in public, reframing XP practices as a 'trust factory' and arguing that as code generation gets cheap, trust becomes the scarce thing.

Christoph Nakazawa on which engineering values survive when coding is no longer the bottleneck, backed by public 90-100% AI-written projects: strong ownership, taste, strict guardrails plus fast feedback loops, context-in-the-repo, own your stack, option value. The capstone for the batch; it reinforces nearly every through-line and the unifying question.

A structural-economic analysis of Google's $80B equity raise and the Berkshire Hathaway deal, arguing cash capacity becomes the ultimate scarce commodity in the AI compute race; uses the See's Candies and BNSF framing to ask which business model wins when compute is the bottleneck.

A practitioner retrospective from Nadella on Microsoft's strategy through the AI platform shift: where margins and moats land, the per-seat vs. consumption business-model shift, agentic coding (GitHub Copilot, Cowork, multi-model harnesses), and the striking claim that a firm's private evals are its most important IP and its tacit knowledge its only real moat.

A practitioner mental model for measuring AI effectiveness as leverage (output per unit of human input) rather than usage; the negative/low/high/maximum-leverage ladder turns on 'how much context does a human need to provide before the AI can do the job well?' Also includes Anna Shipman's 'bottom line up front' exec-translation tip.

Benedict Evans is a structural analyst who explicitly resists hype ('AI is as big a deal as the internet, and ONLY as big'); he reframes the job question as 'is this a task or a job?', argues distribution becomes the moat as software gets cheap, and explains the surprising consulting boom. A podcast, so the substance is in the audio.

A practitioner 'existence proof' that AI changes the calculus of what's worth building: small, weird, mostly audience-of-one tools (offline plant ID, auto-generated Anki cards, an endless wiki) that would not have existed otherwise.

A non-technical recruiter ships a production App Store app via a Claude-as-architect, Claude-Code-as-engineer, Terminal-as-executor loop, a concrete data point on who gets to build software when execution cost collapses and how technical expertise is being redefined.

A designer's end-to-end practitioner retrospective on building a personal tool (audience of one) with Claude Code, documented warts and all: 429 rate limits, masonry and pagination tradeoffs, Git worktrees for side-by-side design comparison, and Vercel lockfile debugging. A concrete real AI build that doubles as an agentic dev workflow.

Kent Beck interviews Randy Shoup (distributed-systems veteran) on the need to make things vs. merely document them, why deep distributed-systems understanding isn't threatened by AI, and how Jevons paradox explains cheap cognition.

Charity Majors on the 'yawning chasm' between AI enthusiasts and skeptics in high-discipline teams, where wins and costs land on different people so there's no feedback loop.

A solo dev documenting a real AI product build (Tolaria) with metrics: 150K LOC plus 100K LOC tests, ~28 commits/day, 99.1% crash-free, ~2 hrs/day, monthly cost down more than 90% after moving Claude to Codex.

A practitioner reflection inverting the usual causation: the best programmers weren't fast because they were great, they were great because they were fast (fast = more data, faster learning, more approaches tried).

Claire Vo's six-part framework for writing autonomous Codex Goals (outcome, verification, constraints, boundaries, iteration policy, stopping conditions) reframes agent use from babysitting a tool to managing a colleague, and maps the agentic loop onto OKR-style outcome definition. No standalone article link; it lives in the podcast and newsletter post.

Argues the blocker on AI velocity is org structure and trust, not tooling: titles become permission structures that gate context, so high-agency employees shrivel.

Camille Fournier on how individual AI productivity can degrade team productivity: don't ask someone to review what you haven't reviewed yourself (the 'validation tax'), shorter is better, AI is not an excuse to turn off your brain or your heart. Crosses the care-for-teammates curiosity with AI-at-team-scale.

Thorsten Ball argues that building new software IS learning, so 'that's not what I meant' is unavoidable; the highest-leverage move is shrinking the time from 'let me try' to reality's feedback via prototypes, fake demo videos, README-first APIs. It hits how planning artifacts and feedback loops shape what gets built.

Molly Graham (ex-Facebook) on memorability as the real test of goals and values: if people can't hold it in their head, they can't use it to decide.

A 35-year practitioner retrospective arguing technical interviewing is 'bordering on pseudoscience,' with a memorable proof (a Google hiring committee tricked into rejecting 2/3 of its own members' packets).

A cross-domain structural analysis (economics, org design, history) using Milgrom-Roberts 'bundles of complementary practices' to explain why Japanese firms diversify and excel at high-precision manufacturing, and why piecemeal copying (the andon cord, performance pay) fails. A standout for connecting unrelated fields.

A three-decade self-taught practitioner retrospective.