What we're reading this week

a rare research synthesis where five independent papers converge on one story: AI compresses code generation while the real bottleneck shifts downstream to verification and shared understanding. A practitioner openly changing his mind.

twelve warts-and-all points from the person who coined "context engineering": the dumb zone, trajectory poisoning, intentional compaction, and a system they threw out after four months of shipping unread code.

agentic coding produces the outputs of writing code without the cognitive processes that build skill; Boykis catalogs concrete friction-adding practices to rebuild your own foundation.

traces loop engineering from Huntley's "Ralph Wiggum" loop to its productization, holding both the real use cases and the disappointment ("tokenmaxxing," agent drift) that it may be a temporary hack.

the creator of Odin on how tools should disappear into cognition rather than become identity or a puzzle-game; the feeling-productive vs. being-productive gap, examined through his own config-fiddling past.

Cal Newport uses Ong's Orality and Literacy to argue literacy isn't merely a tool but the technology that produced logic and the independent self, lifting it above the "reading is dying" framing it answers.

Armin Ronacher uses the Tower of Babel as a lens on how agents strip away the coordination friction that used to synchronize human understanding of a codebase. The thematic capstone of the week. (Surfaced by both Hacker Newsletter and Matter.)

Narayanan & Kapoor use a "decide-execute-deliver sandwich" model plus labor economics and forensic debunking of "AI-washed" layoffs to argue value migrates upward to irreducible decision-making.

useful software design is concrete (individual files and lines) and can't be done by detached architects; the line between translating generic advice and actually understanding a system.

reframes politics not as scheming but as reading how the organization actually operates: the tacit, undocumentable capability separating effective engineers from invisible ones.

Maggie Appleton's attempt to define the "design engineer": knowing the medium's constraints collapses the artifact/handoff chain so one person can hold design-to-implementation in their head.

Bun's 535K-line Zig-to-Rust rewrite done in 11 days with 64 parallel agents (git-stash collisions, 16K compiler errors, $165K in tokens): AI makes previously-impractical migrations feasible, changing the economics of what gets built.

Claire Vo's walkthrough of a real Sentry-to-Linear triage harness demystifies "harness" concretely (tool adapters, permissions-as-flags, structured artifact bundles) and argues owning the harness layer is the durable moat.

treats the compute-to-payroll ratio as the new dominant cost of software engineering (Anthropic at 2.3x payroll), triangulating sources into three 2029 scenarios. Economics, not opinion.

a first-principles, cross-domain argument (information theory + stylometry + demographics) that as de-anonymization costs collapse, privacy and pseudonymity become the scarce things, and VPNs and pen names won't save you.

a teardown of value capture when creation costs collapse: Canva makes pro tools free, turning Affinity's power users into a loss leader to monetize the far larger pool of office "normies."

an agenda-free retrospective on craft over scalability: the toolmaker's dilemma, building Ghostty for an audience of one, and why understanding computers matters more than any language.

a solo builder's concrete architecture for a 24/7 local AI fleet (Mac Studio, DGX Spark, RTX 5090) routing cheap continuous local grunt-work to frontier models for judgment. Self-hosting, build/review loops, and small-team leverage at once.

a hands-on build of a fully local coding-agent stack (Ollama + open-weight models on a Mac Mini / DGX Spark wired into Qwen-Code, Codex, Claude Code) with custom benchmarks and a non-obvious finding about where the token cost really goes.

Birgitta Böckeler (Thoughtworks) structurally decomposes what actually gates local models for agentic coding (RAM, MoE architecture, quantization, tool-call schema mismatches) rather than hype.

Will Larson maps teams onto four states (falling behind, treading water, repaying debt, innovating), argues hiring is over-reached-for, and prescribes a distinct lever per state plus consolidating resources on one team at a time.

a real Code Yellow at Provet becomes a study in how shared vocabulary and a reusable template (problem statement, exit criteria, "tap on the shoulder" authority) mobilize a whole org under pressure.

Honeycomb's CTO reframes the reviled AI mandate not as coercion but as a funding mechanism that forces orgs to name tradeoffs and allocate real time for skill-building.

recasts resistance as information, not a character flaw, with a six-profile readiness taxonomy and the "immunity to change" lens; invest in the persuadable middle, not the loudest deniers.

Kent Beck on how ideas ferment over years and, more sharply, on the mechanics of explanation: refining an idea by explaining it badly, repeatedly.