Newsletter #28 • November 2026

Building with AI

• Agentic Verification → Why agent quality comes from verification loops (e2e tests, property-based testing, benchmarks) rather than humans reviewing every line
• Are Harnesses Worth the Cost? → A 10-run experiment where a behavioral harness cost 7x more and used 10x the tokens without improving GPT-6 Sol’s results
• Harnais agentiques (1/2) : pourquoi les bonnes pratiques d’hier coûtent cher aujourd’hui → Pourquoi les harnais conçus pour des modèles fragiles deviennent inutiles, puis contre-productifs, face aux modèles actuels et à la nouvelle économie des tokens
• How Kent Ships Features Without Reading the Code → Kent C. Dodds’ fifteen principles for shipping agent-written code he hasn’t read, from intent to gates to feature flags, with the session recording
• Plan mode is dead → Why plan modes no longer help agents but understanding what you build matters more than ever, from the founder whose planning-first app failed
• Self-improving software is inevitable → Why the next step is closing the loop: feeding production signals back to agents automatically so software fixes itself
• Five months treating bugs like patients and coding agents like a medical team → How Cockroach Labs’ AI “teaching hospital” shipped IBM Db2 migration support in under two days: 164x faster and 38x cheaper than the earlier Oracle migration work

Code Review & AI

• How the Claude Code team does code review → How the Claude Code team doubled merged PRs while cutting reverts by 40%, moving most code reading to machines and making mistakes easy to undo
• We stopped reviewing code and started verifying intent → Why Tessl dropped line-by-line code review for a verification layer built around intent after PR volume quadrupled
• How to read code → Why reading code isn’t like reading prose, and how to read diffs and unfamiliar programs effectively

AI Ecosystem & Skills

• OpenCode 2.0 → A shared background server for TUI, web and desktop, Code Mode for running MCP servers as scripts, and background subagents
• mattpocock/skills releases → Matt Pocock’s agent skills for real engineers, now with a code-review skill that enforces your repo’s standards and a dated LLM-wiki memory
• break-ui skill by Emil Kowalski → An agent skill that stress-tests UI with worst-case data (long names, emoji, empty states, huge counts) and reports every break with a fix

Models & Research

• Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents → Meta’s new agent-focused models, with Active Memory across conversations and Active Focus to persist through hard tasks and recover from errors
• Introducing Clairic, a new System One class of extraction models → Cloudflare’s extraction model returns typed values with confidence intervals in one forward pass, halving cost and time versus an all-LLM pipeline
• TokenEconomizer: Experience-Driven Black-Box Cost Control for Coding Agents → Microsoft Research and UC Berkeley’s black-box method to control coding agents’ token costs using experience from past runs
• Problems in 22 scientific fields had solutions hiding in plain sight. An AI has found them → How the BootLoops AI harness borrowed methods across disciplines to produce 36 papers in weeks, from linguistics to health policy
• The Mathocalypse → Scott Aaronson on AI agents cracking open math, including Claude Code proving BusyBeaver(10) exceeds Graham’s number

AI & People & Process

• The Manager’s Path in the Age of AI → Camille Fournier’s LDX3 talk on what AI really changes for engineering managers, against CEOs declaring management dead
• How PM Work Is Changing in the AI Era → Interviews with PMs from startups to public companies on how AI became the context layer for their daily work
• What Do You Still Need to Know When Agents Do the Work? → What engineers still need when agents write the code: fundamentals, breadth, and the willingness to jump in when it breaks
• When code is cheap, judgement becomes the job → How a 21-person team ships hundreds of PRs a week with agents, auto-merging 16% of them

Trends

• Code Is a Commodity Now → Job-market and productivity data on why producing software is deflating, and six predictions from team compression to specs replacing codebases
• The state of devs 2026 → 5,463 developers on burnout, AI, layoffs and discrimination: 62% report burnout, and nearly half say AI hasn’t helped their mental health
• The people holding up the internet → A data-driven look at the handful of volunteers maintaining the software under billions of phones, like the time zone database

Learning From the Field

• Cutting our inference bill by 68% by being obsessive about optimization → How Cloudflare cut its AI code reviewer’s cost per run by 68% with prompt caching and model routing, while making reviews 2.7x faster
• Node.js at scale, rebalanced: how we cut cost by 38% → How Booking.com cut its biggest Node.js service’s compute costs by 38% with platform and config changes alone: 30% fewer pods, 20% less memory
• How I made our CI 20x faster (steal this) → How Boardy cut CI wait from 40 minutes to 2 with faster runners, Oxlint, TypeScript 7 and more, so agents stop waiting
• CI Playbooks: prompts and skills for fast CI → Agent-ready GitHub Actions playbooks distilled from how Anthropic, Stripe, Shopify, Uber and others speed up CI
• Redesigning source control at Anthropic → Why Anthropic built Scion, an in-house version control system, after merges grew from 2,500 to 30,000 per day, and migrated in five months
• Hopper: we built an internal coding agent, now 80% of Alan uses it → How Alan grew a codebase Q&A bot into one shared agent for the whole company, behind a single MCP gateway, now opening 25% of merged pull requests

Security

• How I Found a $90,000 Remote Code Execution Vulnerability in Google → A forgeable Flask session cookie plus pickle.loads gave unauthenticated RCE on developers.google.com and four other Google sites
• Coordinated disclosure of critical vulnerabilities → How OpenAI led a coalition to patch critical PHP vulnerabilities found with GPT-5.6 Sol before attackers could exploit them
• The Secure Software Factory: A Reference Architecture → A reference architecture for agent software factories: the agent proposes, policy authorizes, infrastructure enforces, across sandbox, network, supply chain and identity

Architecture

• The ‘generic’ service circular dependency problem → How to stop generic notification, analytics or comment services from accumulating every caller’s types and creating circular dependencies
• PostgreSQL uuidv7 vs UUIDv4 vs bigint keys, measured → 10 million rows on PostgreSQL 18: uuidv7 keeps pace with bigint, while UUIDv4 doubles load time and triples WAL
• Introducing @nestjs/resilience → Retries, timeouts, circuit breakers, bulkheads and fallbacks as NestJS decorators, with idempotency-aware retry defaults out of the box

React / React Native

• TanStack Start 1.0 is Stable → TanStack Start goes stable with typed server functions, selective SSR, full-stack middleware, and a formal API stability commitment
• Optimizing the sh*t out of TanStack Start → How cutting downloads and fixing fonts and images took a TanStack Start home page from 73 to 100 on mobile Lighthouse

Tooling & Ecosystem

• Introducing tsrs: a Port of TypeScript to Rust → Cloudflare and VoidZero port the TypeScript compiler to Rust inside Oxc: up to 8x faster and 70% less memory than TypeScript 7.0
• Ghostty 1.4.0 → Full Kitty graphics support, 70-90% less scrollback memory, a new Linux OpenGL renderer and a ghostty +ssh wrapper that installs terminfo remotely
• Hex → Local voice dictation for Mac, Windows and Linux: hold a key, speak, and it pastes, with nothing leaving your machine
• Learn Arduino → Interactive 3D Arduino lessons for beginners, with a pinout explorer, component library and built-in simulator in the browser

News

• Nvidia débarque au capital de Doctolib → Nvidia prend une participation minoritaire dans Doctolib, valorisée 3,6 Md€ avec 422 M€ d’ARR, et investit aussi dans Pigment

Funny

• Serverless Horrors → A collection of real serverless billing nightmares, from a $10K Durable Object loop on Cloudflare to a $100K Firebase bill
• Hacking Cheap Smart Photo Frames into a Family Dashboard → How Wes Bos rooted a $50 Android photo frame into a family dashboard, fed by an agent that turns school emails into events