Rangle

What we're reading this week

On November 13, a global Microsoft outage and a nationwide FAA ground stop hit on the same day. Why does software keep failing like this? Jimmy Breck-McKye answers with one word, decoherence: changes that each make sense alone and stop making sense together, arriving faster now that agents write most of them. He thinks companies risk losing the ability to think and respond as one unit. That seems like the kind of thing an owner can take care of, and it shows a role that still needs to be filled. Recently one of our agents reverted a senior developer's deliberate change because the docs didn't cover it. Each step made sense alone, and it was caught only because the agent stopped and asked a person.

Ben Hofferber
Curated by Ben Hofferber

How to read this

📰 Quick hits

The headlines worth knowing even if you read nothing else this week.

Fable 5.2 released; reviewers position it alongside Opus 5.5

Fable 5.2 is out across the API, apps and Claude Code. Claire Vo finds it fast and creative with strong spatial reasoning, but weak at unguided visual design, and keeps Opus 5.5 as the default for most coding.

via Lenny's Newsletter (How I AI)

Anthropic IPO raises $107B, the largest ever

Anthropic debuted on Nasdaq after raising $107B in its IPO, five and a half years after founding.

via The Pragmatic Engineer / Hacker Newsletter

Google releases Gemini 4 Neon

New Gemini 4 model; Microsoft also announced seven MAI models at Ignite.

via Hacker Newsletter

Standalone Claude Design closes December 14; design MCP and /design-sync stop working

Design now lives in regular Claude conversations. Projects on claude.ai/design must be moved or downloaded before Dec 14, and chats and comments don't migrate. The standalone MCP server, /design-login and /design-sync stop working, while /design in Claude Code is unaffected.

via Claude Team

Harvard study: AI coding agents moved the bottleneck to code review

The Cowork, Harjot Gill and Devon Estes pieces in this issue each describe a team working around that bottleneck.

via Hacker Newsletter

Costco lets AI autonomously reorder pharmacy drugs; 90% of retailers limit agents to look-up access

Costco's pharmacy AI compares vendor prices and places reorders itself (in-stock above 98%), while a PYMNTS survey finds 90% of retail firms would restrict AI agents to look-up access and none expect full autonomy within five years.

via Retail Dive: Tech Weekly

🧠 How tools reshape cognition

The questionWhat does using AI actually do to how we think, learn, and pay attention?

Engineer Role

Software is breaking down at a systemic level

Jimmy Breck-McKye · breck-mckye.com

Breck-McKye traces 2026's outages through cloud, open-source dependency chains and the flattening of specialists into generalists, and ends at AI-driven "decoherence": changes that are each valid locally pile up into systems nobody understands, while companies lose the ability to think as a unit.

"I believe that tech companies risk losing their 'organisational cognition' - their ability to think and respond as a coherent unit."
Harness Design

Claude Code and Obsidian: Scott Farrar's Household Brain

How I AI

A real Claude Code plus Obsidian household knowledge base, warts included: a linter to stop agent bloat, bad facts persisting in memory files, and sync compromises with a spouse. The design principle is that the AI prepares and the human decides, and the author's tutoring stance has AI help kids think rather than think for them. The author's fuller write-up is at sfxmath.com/work/homeai-household-knowledge-base/.

"I wanted to build a system where I was confident that AI was grounded in my truth."

💰 Value concentration when creation costs collapse

The questionWhen building something gets cheap, where does the value (and the money) actually pool up?

Reviewing AI CodeEngineer Role

Bad Ideas That Aren't Bad Anymore

Devon Estes · devonestes.com

An engineering leader with about 15 years in software changes his mind in public: no peer PR review, one person per work stream, and "don't go deep" career advice all work now. He traces each old rule to a cost that AI removed and offers a test for checking your intuition.

"What's the cost that made this a bad idea, and is that cost still there?"
Reviewing AI Code

Software factories and code reviews with Harjot Gill

Gergely Orosz · The Pragmatic Engineer

An extreme counterpoint to the Cowork piece: CodeRabbit's CEO ships 5,000 PRs a month on about $250K a month of AI spend, built his own CI after a $13M/year quote, reviews releases instead of PRs, revokes review exemptions after outage "strikes", and hires for strategy-game hours. It is useful to weigh against the more careful approaches.

EvalsHarness Design

Software engineering after code becomes cheap with Satya Nadella

Gergely Orosz · The Pragmatic Engineer

Nadella admits the IDE faded in six months and caught Microsoft off guard; he frames verification as the unsolved problem now that generation is cheap, argues for owning the harness, memory and model with private evals, and calls decision models "the stored procedure of the intelligence era".

Obituaries

Ben Thompson · Stratechery

Structural analysis of Nvidia's moats after Huang's exit: LLMs dry up software moats like CUDA, leaving supply-chain trust built on one man's handshakes and a company organized as an exoskeleton around its founder (60 direct reports, no succession), which ties to small-team leverage and its fragility.

"LLMs are so effective at writing software that any software-based moat will inevitably dry up."

Token Fraud Is Eating AI Margins. Here's How Claire Vo Is Fighting It.

How I AI

A solo founder's costed retrospective: tens of thousands of dollars lost to trial farming, promo-code resale and MCP compute abuse led her to kill free trials, cap spend in dollars and bound her support agent. It shows structurally why the free-trial model breaks when every request costs real compute. It is a podcast recap, so specifics like thresholds are absent.

"If you sell AI, you sell compute."

🚀 Small teams, disproportionate output

The questionHow do tiny teams punch so far above their weight?

Harness DesignReviewing AI Code

Nine engineers, 1,211 pull requests in a day: Lessons from building Copilot Cowork

Microsoft Command Line

Detailed practitioner account of a nine-person team going from an empty repo to GA in 4.5 months and a live Python-to-Rust migration with agents writing most of the code. Lessons include no code without a spec, checks that must prove they checked something, rules kept in the repo, a crate graph shaped for parallel agents, and loops given targets and limits.

"Agents keep working until the checks pass, so if a check can pass without testing anything, an agent will eventually find that path."

🤝 Intentional hospitality as a practice

The questionWhat if you designed care into how you treat people, deliberately, as a real competitive (and moral) advantage?

The Conversation Before the Conversation

Kevin Goldsmith · blog.kevingoldsmith.com

Kevin Goldsmith looks back on losing engineers at Adobe, Spotify and a CTO job by postponing easy conversations, and gives five signals that one is overdue (pattern, distance, workarounds, cognitive load, inflation); the cost of waiting falls on the person who never got the feedback.

"I used to think I was being patient. Most of the time, I was just being slow."

⚙️ Agentic development patterns

The questionWhat do the setups, protocols, and loops that actually run coding agents look like in practice?

Canadian connection: Toronto headquartered; published by Rangle.
Harness DesignReviewing AI Code

When the code and the docs disagree, which is the source of truth?

Ken Easson · Rangle

Ken Easson from our team on the week one of our agents reverted a team lead's deliberate change because the behavior documents didn't cover it, then stopped and asked. He sets out the "needs decision" lane we use, so a person says whether the code or the docs get fixed.

"If the most senior people on the team make deliberate changes to the code which violate the behavior, it's the behavior that's poorly defined."
Canadian connection: Canadian author.
Harness DesignScaling Yourself

Guidelines and Guardrails Belong in Git

Allen Pike · allenpike.com

Pike argues that the rules a team's agents follow belong in the repo, shared and owned, and that a deterministic check beats a prompt wherever one is practical. He notes teams that review every code change but not the AGENTS.md file every agent reads, and a "start with a failing test" rule that left one codebase with thousands of useless tests.

"Basically, in-repo rules turn expertise into leverage."

👀 On the radar

Lower-confidence picks worth a skim if the topic grabs you.

  • How to (actually) get promoted Practitioner who has sat in hundreds of calibration meetings names four factors beyond good work (legibility, likability, location, leverage); the claim that likability is substance and the note that context becomes the scarce thing above ~50% headcount growth are the parts beyond generic career advice.
  • Science isn't a series of lucky accidents Saloni Dattani debunks the Fleming, Pasteur and Rontgen accident myths and argues discoveries came from systematic trial and error ('engineered serendipity'); well sourced but mostly history of science with a loose fit to the hard-slow-work theme.
  • Why no one wants you to click on a retail website anymore Short reported piece on agentic checkout (Google Universal Cart, Meta Muse, Shopify agents, Alexa for Shopping) that frames the shift as retailers moving from winning the click to being chosen on platforms they don't own, with Amazon walling off its ad business; mostly analyst quotes, little depth.
  • Meta's Muse: How we launched the most successful consumer product this year | Mona Sarantakos Podcast with Muse's product lead on a real AI consumer build (an OpenClaw-on-Raspberry-Pi weekend that grew into Muse, a seven-month wait before launch, agent swarms for user research); the web post holds only the topic list, so the substance is in the episode.
  • Coordination, old intuitions, and readings! Luca separates collaboration from coordination (sync, async, none), traces most coordination to cognitive load, and argues AI lets people merge specialist layers again; a second note says our sense of what is hard is muscle memory from the hand-built era.
  • Introducing Review Box Radoslav Stankov built his own PR review app around how he reviews (1,197 reviews this year); now that agents write the code, he leaves bugs to AI reviewers and spends his own review on requirements, architecture and how people's work fits together. It is partly a product launch.
  • What will the next generation of developers do when AI writes the code? (PRAISE) Microsoft and CMU borrow the preceptorship model from nursing to pair junior engineers with seniors who pass on taste and judgment about AI use; the idea is interesting but the piece is short and partly an announcement.
  • How AI Will Change Engineering Leadership In 2027 A grounded engineering-leader view: most 'AI-native' products could always be imagined but cost too much labor, so start with a human-in-the-loop workflow and swap in AI; prefer prototypes to designs; the models are good enough; and two rules against slop. Useful, though it mostly confirms the consensus.
  • Asking Better Questions A short piece on giving one sentence of context with a question; useful context changes the urgency, the format of the answer, or the definition of done. It applies the idea of a small courtesy that removes anxiety.
  • Inference engineering for software engineers Structural look at why companies leave frontier APIs for open models on inference providers (Uber -50%, Pinterest -90% per-token), the path from multi-tenant APIs to dedicated GPUs, and private evals as IP; mostly infrastructure detail.
  • For us, AI is finally delivering Luca describes his four-person team's org chart of bots (a chief of staff for each person, area managers, PM bots that build internal tools, a shared vault, retros run by a 'biographer' bot); the eight collaboration patterns are behind the paywall, so the free text only outlines them.
  • My Least Favorite Example in Software Engineering Short essay on what makes a teaching example good (it shows why the topic matters, not only what it is), traced through the history of the double-reverse property-based testing example; it speaks to how understanding transfers.
  • The Pulse: What can we learn from the Railway hack? First detailed account of an attack driven by an abliterated (safeguard-free) model: stolen source was weaponized within hours, and defenders were limited by models that refuse security work. Also contrasts Railway's CEO helping a rival with Northflank's opportunism, which then backfired.
  • We Need Smart People Who Give a Shit Argues that fragile systems like ACH run on people who combine intimate knowledge with care, and that replacing their work with prompting erodes both. It shares a thesis with the systemic-breakdown essay but reads more as opinion than analysis.
  • AI is killing All About Berlin An independent publisher lost 75% of his traffic to Google AI Overviews after seven years of growth; it is a concrete case of value moving from the people who make knowledge to whoever aggregates it. The post is from May and has a follow-up update.
  • Design For Understandability Google talk arguing that understandability should become a design norm the way testability did, using contracts, invariants and isolated side effects for local reasoning; without understanding, LLMs act as complexity generators. The fundamentals are old, but the framing is current.

How this is made

Each week, Ben works through 32 newsletter emails from 15 publications, and the 11 pieces worth your time land here, distilled into about a 6-minute read. AI helps surface and summarize the strongest pieces; they are grouped by the question each one is really wrestling with, and anything that connects to a practice we are building at Rangle gets a tag. Every pick, summary, and tag is reviewed by hand before publishing.