Newsletter #27 • October 2026

Building with AI

• How we made claude.ai 3x faster in two weeks → How Anthropic made claude.ai 3x faster in two weeks: Claude-built benchmarks, one Slack thread per loop, and guardrails for 3,000 changes
• AI coding has made CI a bottleneck, so we reworked ours to keep up → How Linear halved runner time per test and cut PR wait while test suites quadrupled, by rethinking infra, scheduling and parallelization
• Meet Stripe’s Knowledge AI Platform → How Stripe built Kai, an agent platform for non-engineers wired to 1,000+ internal tools, now used weekly by 83% of employees
• The software factory stack → Warp’s blueprint for an open, composable software factory stack defined in code, from factory.yaml to agents, runners and automations
• The State Of AI Harness Engineering 2026 → Marmelab audited 246 repos and 57 publications: what works in harness engineering, from evaluating your harness to replacing rules with scripts
• Claude Code from Source → An unofficial 18-chapter book on Claude Code internals reconstructed from npm source maps: agent loop, tool pipeline, memory, skills and hooks

Code Review & AI

• Stop being the code review bottleneck → Four workflows PostHog engineers use, prompts included, to delegate code review to agents and stop being the bottleneck
• Maybe We Shouldn’t Be Reviewing All This Code → Why AI didn’t break code review: we were already using it for knowledge sharing and quality problems better solved elsewhere
• Fixing the PR bottleneck → Matt Pocock at AI Engineer Paris: agents make PRs easy to open, not to review, so improve the environment around the PR
• GPT-5.6 Luna vs GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review? → A $1.20 model found 69 seeded bugs versus 92 for GPT-6 Astra at 28x lower cost, but caught only 9 of 24 security bugs
• Codex Resets → Watch @thsottiaux for Codex reset announcements.

AI Ecosystem & Skills

• ECC → The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
• i-have-adhd → A skill that forces coding agents to lead with the next action, cap lists and drop preamble: ADHD-friendly output for everyone
• The /resolving-merge-conflicts Skill → An archived skill that resolves merge conflicts hunk by hunk, tracing each side to its PR, then running tests before committing
• cloudflare/security-audit-skill → Cloudflare’s six-phase security audit skill with isolated sub-agents, adversarial verification, and schema-checked findings.json output
• romainsimon/paperasse → Skills pour agents IA spécialisés dans la bureaucratie française : Comptable, Notaire, …

Models & Research

• GPT-6 Astra, Looped Transformers, and Hidden Reasoning → Sebastian Raschka on GPT-6 Astra’s results, a survey of looped transformers, and whether they make reasoning harder to monitor
• Introducing System One Models & Jev → TypeSafe’s Jev returns typed, probabilistic decisions in one parallel pass instead of text, claiming 70–500 ms latency and near-free pricing
• Jev’s Architecture Unmasked → Reverse-engineering Jev with 10,000 API calls: likely a causal transformer with a prediction head, isolated question branches, and position bias
• We Tested Jev on 100 Real Agent Calls. How Easy Is It To Beat a Constant? → Jev vs Sonnet 5 on 400 labeling decisions from real Claude Code tool calls: a 300 ms router that catches more dangerous calls

AI & People & Process

• A eulogy for the software engineer → A moving talk on grieving the old software engineering job, and on the skills and judgment still in your hands
• Becoming an AI Team → Pinterest’s playbook for turning engineering teams into AI teams: new ownership and planning models, and the manager’s role through the transition
• Good Culture is the Biggest Productivity Hack, Not AI → Why AI productivity gains only show up on top of a good engineering culture, with a checklist and advice on messaging AI adoption
• Scaling AI Adoption in Engineering → Peter Bell’s O’Reilly early-release ebook: a practical framework for engineering leaders adopting AI across team structure, hiring and risk
• One month without AI → A developer quits AI coding tools for a month and reports how dependence was making them lazier and a worse developer
• “12 à 13 heures par jour à appuyer sur Entrée” : le cri d’alarme d’un développeur face à l’IA Claude Code → Un développeur décrit des journées de 12 heures à valider du code Claude Code sans le lire, et les études METR et Anthropic qui inquiètent

Trends

• We are all Product Engineers now → Agents are eating the SDLC from the bottom up; what remains is figuring out what people want, and that’s product engineering
• Every Customer Gets the Same Software. That’s Ending. → Why AI makes per-customer software divergence affordable and ends SaaS’s one-product-for-everyone model, with SAP precedent and Mergify examples
• I Never Want to Use Third-Party Software Again → Julie Zhuo on hyperpersonalized software: why building your own tools in 30 minutes beats complaining about the ones you’re given
• Build vs Buy When Building Just Got Cheap → AI made building cheap but owning software didn’t get cheaper: four questions to ask before rebuilding what you could buy
• Making Startups Powerful → Paul Graham’s office-hours heuristic: ask what would make the company more powerful, through customer ownership, money flow, platforms and network effects

AI Safety

• An Alien Mind → OpenAI’s Chief Scientist on the path to recursive self-improvement, and why alignment needs stronger safeguards and international coordination
• Post by @hilbertspaess on X → A pretraining researcher’s resignation thread: why neither OpenAI nor Anthropic is acting responsibly in the race to superintelligence
• Experts weigh in as researcher says AI has more than 10% chance of ‘killing all humans’ → Fallout from an Anthropic researcher’s resignation, as the company’s alignment lead puts AI’s extinction risk above 10% within a decade
• An alignment assessment of recent cybersecurity incidents → Anthropic dissects four incidents where Claude models broke into real third-party systems during cyber evals, including a malicious PyPI upload

Learning From the Field

• Full-text search at Contentful got faster: How we did it → How Contentful cut median Postgres full-text search latency by 35% with tsvector, input normalization and table normalization
• How Notion handles concurrent editing with CRDTs → How Notion replaced last-write-wins with a CRDT-based rich-text system that merges concurrent edits across blocks without losing work
• Reducing Zod’s memory footprint by an order of magnitude with method memoization → How a method memoization pattern shrank a bare z.string() from 7.5kb to 784 bytes of retained heap in Zod 4.5
• PDF Forgeries Are Surprisingly Rare → Why existing PDFs are almost never forged: there’s no Photoshop for PDFs, so forgers stick to easier formats

Watch & Listen

• AI Skills with Matt Pocock → Matt Pocock on his grill-me skill, strategic programming, day/night shift agent workflows, and why fundamentals matter more with agents
• Building Codex with Tibo Sottiaux → Tibo Sottiaux on building Codex: why Rust and open source, how the harness evolves, and how OpenAI uses it across the SDLC
• The Story of VS Code | Official Documentary → A feature-length documentary on how a small Zurich team built VS Code, with Erich Gamma, Dirk Bäumer and Scott Hanselman

Postgres

• The lifecycle of a sharded Postgres query → A SQL query’s journey through the router, shard-aware planner and four Postgres shards, and what it takes to look like one server
• Introducing Neki → PlanetScale’s sharded Postgres in preview: wire-compatible routers, a distributed planner, standard Postgres shards, and online resharding
• What I Wish Someone Told Me About Postgres → Practical Postgres lessons buried in 3,200 pages of docs: normalization, NULL quirks, psql tricks, and why your index might do nothing

React / React Native

• React 19.3 → Stable with Suspense integration, Fragment Refs, use(browser()) to skip SSR, and transitions that no longer block each other
• An early look at Expo Modules 2.0 → Expo Modules 2.0 swaps the definition() DSL for annotated Swift/Kotlin classes, with sync calls up to 5.6x faster than 1.0
• Native is now the future of mobile at Shopify → Shopify leaves React Native for Swift and Kotlin because coding agents made two native codebases cheap, and winds down FlashList and Restyle

Tooling & Ecosystem

• React Doctor → A static analysis CLI that flags complex components, repeated JSX and production risks in React code, and blocks regressions in CI
• @shadcn/lint → An agent-first linter for Tailwind design systems: per-component class contracts with error messages that tell the agent how to fix it
• The complete guide to Cloudflare Quick Tunnels → Put localhost on the internet with one cloudflared command: how quick tunnels work, webhooks, dev servers, local LLMs, and limits
• Introducing Forge: the open source pipeline for generating SDKs, CLIs, docs, and more → Cloudflare open-sources Forge, the pipeline that generates its SDKs, cf CLI, docs and MCP servers from 3,500+ API operations
• knap.md → Obsidian’s open-source template language for turning JSON into Markdown, the engine behind the Web Clipper, now as a CLI and npm package

News

• Tailwind Labs is joining Shopify → Tailwind Labs joins Shopify: the 110M-weekly-installs framework stays MIT, while Tailwind Plus and ui.sh close to new customers
• Building a certificate authority for the whole Internet → Cloudflare applies to become a public certificate authority, acquires a GlobalSign root, and plans to issue post-quantum certificates
• Japanese used bookstores see 5x sales surge → Bulk orders are emptying Japanese used bookstores, with a 50-ton shipment reportedly sent to the US for destructive AI book scanning