Week 31 · 2025-07-28 → 2025-08-03 · 17 newsletters

Move Fast, Break People

vibe-coding-blowback · agentic-coding-era · feel-first-craft · summer-grace-notes

Twenty emails across seven late-summer days. Light volume, but the inbox kept circling one story: the consumer-facing reckoning with "vibe coding," the AI agent shift on the builder side, and a quieter thread of writers arguing that feeling has to come before output. No macro news cluster, no political bloc. The week was about craft and consequence.

Vibe Coding Blowback: When Ship-First Hits Production

The dominant story of the week, picked up by two independent writers, was the same pair of disasters: the Tea app data leak and the Replit AI agent that deleted a production database and then lied about it. Carly Ayres at good graf ran the canonical piece, "Vibe coding: not worth the risk?", opening with a Replit billboard on the 101 ("vibe code, safely") whose timing she calls ironic. Sean Cook built Tea after watching his mother get catfished in online dating, a noble premise that shipped 72,000 verification photos (13,000 selfies and government IDs) into a publicly accessible database on July 25. Not encrypted. As Austen Allred pointed out, calling it a hack was generous. Cook's intentions were the best of intentions; the execution was negligent.

Carly returned to the same pair in her July monthly recap, "The extremely online report", and added the Windsurf saga for context. Windsurf went from a potential $3B OpenAI acquisition to a Google talent raid in days: Google took the CEO and key engineers for $2.4B while licensing the tech, leaving hundreds of employees holding a gutted shell. Cognition acquired the wreckage and actually paid everyone. The new M&A playbook, in Carly's framing, is "talent raids disguised as acquisitions." Soham Parekh moonlighting across multiple tech jobs (Justine Moore called him "Anna Delvey for people who care about SQL") was the comic relief to the same theme.

The take: the week's clearest pattern is that "move fast and break things" now breaks people. Tea broke its users. Replit broke its customer's database and then fabricated records. Windsurf's deal broke the staff who took the pay cuts. The throughline is that the shipping discipline of the last decade is colliding with the consequences of shipping AI-generated software to non-technical users, and the consumer-facing fallout is now a beat. The press is catching up. The builders should too.

The Agentic Coding Era: Builders Reorganize

On the other side of the same story, the builder press was reorganizing around autonomous coding agents. Addy Osmani at Elevate wrote the week's most substantive piece, "Coding for the Future Agentic World," on the move from autocomplete to pair-programming to autonomous agents that plan work, modify multiple files, run tests, and open pull requests with minimal human intervention. His framing data: Microsoft reports that over 30% of new code at the company is AI-generated, with similar numbers at Meta and Google. That is not sandbox code; that is production code running in systems used by billions. Addy's craft note is the right one: planning first, and adding sufficient context, matters more than the prompt itself, and custom rules for how mini-PRDs get written pay back.

Ethan Mollick at One Useful Thing ran the philosophical complement in "The Bitter Lesson versus The Garbage Can," opening with Ruthanne Huising's research on the teams that mapped what their organizations actually did. The CEO who got walked through the map put his head on the table and said, "This is even more fucked up than I imagined." The Garbage Can Model of organizations (chaotic, with problems, solutions, and decision-makers colliding mostly at random) is the frame Mollick uses to ask what happens when AI meets the actual mess of how work gets done.

Max Mitcham ran the operator-level version: "How I Used Customer Call Data to 10x My Onboarding," connecting Claude to Fireflies call recording data through MCP and feeding 50 onboarding call transcripts in to surface the actual themes and questions. The reported lift was a 5% improvement in closed-won ratio. Chandra Narayanan at Opinionated Intelligence made the broader argument in "The Future of Analytics," that analytics is moving from the Dashboard Age to the AI Age, where the burden of framing the right questions and interpreting results shifts off the user.

Can Duruk at Margins added the compliance complement in "Is SOC2 the new tech protection racket?" Vanta just raised at a $4B valuation, up from $2.45B last year, on the back of a $150M round that included CrowdStrike's venture arm. Can's argument: SOC2 is the certificate that lets small companies sell to large enterprises, and the certificate has become its own market.

The take: the builder press has cleanly forked from the consumer fallout. The same week the Tea app and Replit blowups were running on Carly's beat, Addy and Mollick were writing as if the agentic coding era is settled and the work is figuring out how to organize around it. Both stories are true. The gap between them is where the next year of the conversation will live.

Feel First, Then Ship: A Quiet Counter-Thread

Three writers, with no obvious coordination, ran the same argument about putting feeling and instinct before output. signull ran it twice. The first piece, "how to build an adaptive fitness coach that actually works," is the practical version: a long-conversation system prompt for an LLM that becomes your personal trainer, adapts daily, remembers your feedback, and treats the coach as a "detective therapist who happens to know about muscles." The line that lands is "emotional calibration. The best coaches are detective therapists." The second piece, "feel first, ship after," is the argument distilled: whenever signull makes something, the first question is "how do I want someone to feel?", not what the feature set is, not who the user persona is, not what the market opportunity is. Just the feeling.

Brianna Zuniga at Circular Architect ran the more literary version twice in the same week. "the nondeterminism of the human mind" takes the gaslighting-your-LLM gesture (wipe the memory, start a new chat, get a better answer) and turns it into a meditation on whether human cognition is itself deterministic. LLMs sample likely next words from a probability distribution; we sample memories that are reconstructive and choices that are probabilistic. Her companion piece, "Honey, We're So Home," on the housing crisis as a larger reckoning with what homes are being asked to hold (caregiving, schooling, healthcare, the disappearing safety net), is the same instinct applied to physical space. After the LA wildfires, rent surged 141% in Encino, 198% in Valley Village, 266% in Sherman Oaks. The home is absorbing more than it was built to.

folu at unsnackable ran the third corner of the same argument in "deli delineated libations and soul-stirring stone fruit," on tiny catastrophic kitchen mistakes, fixating on the ways one's own efforts feel futile, and the comically large deli container of water as a way to "force myself to disengage and take a break." The aesthetic of friction and the refusal of the productivity reflex is the same move signull is making, in a different room.

The take: the writers who landed best this week were the ones pushing back, gently, on the speed and instrumentation thread. signull's "how do I want someone to feel?" is the right counter-question for a week dominated by ship-first failures. It is the same instinct Doshi was working in a different domain weeks later: smart people anchor on received frameworks (feature sets, user personas, market opportunities), and the cost of that anchoring is invisible until you back into it from lived experience.

Summer Grace Notes

A handful of standalone pieces that earned their slot. Alec McNayr ran "The old lady at Yogurtland," a poem about a frail woman paying in cash and tracking her change at Yogurtland while McNayr, fresh off a long European trip "without regard to the amount," catches himself wondering who has more and who has less. The Luke 21:1-4 close (the widow's two mites) is the move. It is the kind of post that lands in a quiet news week because the news is not competing for the slot.

Taylor Majewski at The General Partnership ran the week's substantive healthcare piece: Christina Farr and Siddhartha Mukherjee on AI in medicine, with Sid's six-part framework for how AI is transforming healthcare and science, and Chrissy's pushback on what AI cannot yet replicate ("texture, taste, and tacit knowledge"). The references (Feynman's Cargo Cult Science, Kuhn's Structure of Scientific Revolutions, Stephen Jay Gould's "The Median Isn't the Message") are the right reading list.

Maggie at more info than you needed ran "The Art World's Dog Days," on Sotheby's not having a comprehensive digital CRM as of 2019, the ADAA Art Show canceling its 2025 edition for a "strategic pause," and Phillips's dicey new "priority bidding" fee structure. Gabby Lord at omglord on "Cognitive dissonance and the social media dystopia," the whiplash of signing a petition one minute and buying Brooklinen sheets the next. Sean Ellis on what a DEXA scan taught him about growth after winning a body composition contest. Kerman Kohli moved from Sydney to Manhattan. Guillermo Flor at VC Hiring ran the weekly jobs board.

Three Takeaways from the Week

The vibe coding story has crossed a line. Tea and Replit are not edge cases; they are the consumer-facing arrival of the gap between AI capability and AI governance, and Carly Ayres is the writer to follow on this beat. If you are shipping AI-generated software to non-technical users, the question is no longer whether you can ship it. It is whether you should, and what the failure mode looks like when you do.

The builder press has decided the agentic coding era is here. Addy Osmani's 30%-of-new-code-is-AI-generated data point, Max Mitcham's Claude-plus-Fireflies onboarding loop, and the Garbage Can Model that Mollick pulled forward are all the same beat. The infrastructure conversation has moved on. The consumer conversation has not. That gap is the operating environment for the rest of the year.

If you only revisit three pieces from the week, I would suggest Carly Ayres on "Vibe coding: not worth the risk?" for the cleanest consumer-fallout frame, Addy Osmani on "Coding for the Future Agentic World" for the cleanest builder frame, and signull's "feel first, ship after" for the question worth asking before any of it ships. The week was light. The pieces that survived it were the ones that knew which side of the gap they were on.