The Harness Is the New Operating System

The Harness Is the New Operating System

Greg Isenberg spent two days in San Francisco with billionaires, data labs and AI founders running one-to-hundred-billion-dollar companies, and came back with twenty-one observations that read less like a trip report and more like an industry constitution. The sharpest one redraws the whole computer: the model is the CPU, the agent harness is the operating system, and skills and plugins are the apps. Everything you thought was a wrapper is infrastructure; everything you thought was the product was the chip.

The implication flips the last two years of snobbery on its head. The “GPT wrapper” sneer assumed the model did the work. Isenberg’s San Francisco consensus says the model contributes roughly twenty percent of the value — the harness, the routing between models per job, the evals, the memory, the A/B testing, does the rest. Vertical agent harnesses are, in his words, the good new wrappers: thin on UI, heavy on APIs, data layers and tools, doing one job’s workflow end to end. The best diligence question for an AI startup is no longer about the model at all: does a new model make you better or worse? If you get stronger when the base model changes, you own the durable layer. If you get commoditised, you were never more than a feature of someone else’s chip.

Two more observations from the twenty-one are worth pinning to the wall. First, value is migrating to data: data brokers are reportedly earning around two million dollars a week selling company data to labs, and legacy SaaS firms may be mispriced — not as software businesses, but as data companies sitting on the training sets of the future. Second, the next fight is memory: remembering the user beats raw model smarts, and real agent-failure logs inside companies are becoming uniquely valuable assets. Click-heavy legacy software doesn’t survive the agentic era; whoever exposes their internal data over MCP first — and almost everyone’s internal data is a mess — wins the integration race. It’s a consultancy-scale opportunity hiding inside every enterprise IT department.

Your saves this week prove the thesis is already productising in three directions. First, Dada’s Product Design Harness at Lumenridge: “STOP LETTING YOUR AI AGENTS BREAK YOUR PRODUCT,” his post opens, and the method is a loop the whole team runs — Document → Prototype → Govern → Build. Anyone on the team can now spin up a working prototype; without a harness, five prototypers hand you five forks of your product. With one, you get five tested ideas and one product — designers governing, engineers merging, the company’s own components doing the building. Second, Instinct, a Y Combinator-backed launch claiming the first frontier agent harness for individual traders: a team from Squarepoint Capital, CERN, the USC Quantum Computing Center and Toss, building agents with their own computers that synthesise market data, news and social in parallel — the organisational capacity of an investment firm, given to one person. “The individual becomes the portfolio manager,” their thread says; edge concentrates in worldview, judgement and bets. Third, Delian Asparouhov’s read on Ramp’s Glass launch: everyone thought Ramp was corporate cards, but they built a Codex/Cowork competitor that’s arguably better — because it’s a third party trying to save you money, not a lab selling you tokens.

The remaining observations are a futures market in miniature. About twelve months from agents that learn your job by watching you work. Unprecedented cloning and reverse-engineering of software with agents — your moat’s source code is legible now. Physical AI rising, with deployment — not research — as the hard part. Cybersecurity, especially post-breach cleanup, heating up because each agent is a new attack surface. Dead startups’ data — Slack, email, code — worth licensing. Monetising the “agent web” still early. Agent insurance coming. The most-wanted hire: someone with taste who writes evals. A new career pattern: raise VC, become a creator, fail, bootstrap as a full-time creator. And the sobering closer: the bar is now a billion dollars of revenue in twelve months — but retention is the hidden risk, and distribution is brutally hard because attention is the scarce resource.

• Own the harness, not the chip. The durable question for any AI startup: does a new model make you better or worse? Better means you own the OS layer.

• Data is the new balance sheet. Brokers selling company data to labs at ~$2M/week; legacy SaaS may be undervalued as data companies; agent-failure logs are proprietary gold.

• Memory beats smarts. The next platform fight is who remembers the user — raw model intelligence is table stakes.

• Prototyping without governance is forking. Lumenridge’s loop — Document, Prototype, Govern, Build — is the template for teams where everyone can build.

• Hire taste that writes evals. The scarcest role in the agentic org isn’t the researcher; it’s the person who can judge what good looks like, repeatedly.

Isenberg’s 21 observations · Dada’s design harness · Aakash Gupta on the chaos without one · Instinct’s trader harness · Delian on Ramp Glass vs the labs

Gemini 4 Argon Week: Launches Are Media Events Now

Sundar Pichai announced Gemini 4 Argon this week — “available globally today for developers and enterprises in the paid tier” — and the launch behaved less like a software release than a premiere. Five point four million views on the announcement post. A thirty-minute video with Sergey Brin and Josh Woodward on how they got here. Elon Musk quote-posting his congratulations with a line that doubles as media strategy: CEOs posting to the public directly on X is the best way to get your message out.

The ecosystem did the rest of the marketing for free. @PromptLLM’s “insane take from Gemini 4 Argon” screenshot pulled 377K views. Pichai himself quote-posted a creator’s video-generation clip, admitting he’d watched it a few times — “especially to the end” — which is the modern launch funnel in one gesture: ship the model, let the internet find the jaw-drop, amplify the best one from the CEO’s own account. It pairs neatly with something David Senra clipped from Alexandr Wang this week: modern product marketing is now all driven by the screenshot. The demo you can repost beats the benchmark you can’t.

The deeper play is the developer surface. Argon ships globally to developers and enterprises first — the paid tier, the API, the tooling — because the model war is fought where the harness gets built. That’s why Logan Kilpatrick’s thirty-minute conversation with Brin matters more than the announcement: it’s the “how we got here and the path ahead” story that tells builders where to place their bets. And in the same week, Sameer Samat announced Google Play is evolving to help people discover “trusted and high-quality apps” — curation as the answer to infinite supply. When anyone can build, the scarce layer is the one that decides what’s worth finding.

• The screenshot is the launch. Wang’s rule, via Senra: the repostable demo beats the un-repostable benchmark — build one jaw-drop moment per release.

• CEOs are the media now. Musk’s read is right: the highest-leverage distribution for a launch is the chief executive’s own account.

• Ship to the harness-builders first. Argon’s developer-and-enterprise-first rollout concedes where the moat forms: whoever routes the model into workflows wins.

• Curation scales with supply. Play Store’s trust push is the platform answer to the wrapper explosion — when everything is buildable, discovery is the product.

Pichai’s Argon announcement · “Insane take from Gemini 4 Argon” · Logan K’s Brin/Woodward conversation · Musk on CEOs posting directly · Wang: marketing is driven by the screenshot

The Voice Wars Go Open

Two voice-AI stories landed this week, and together they describe a market being squeezed from both ends. From the top: FishAudio’s Drama 3, announced with a $52M seed round and 15M+ users, 12M views on the reveal video — the well-funded, closed, productised end of voice. From the bottom: open weights, arriving with a punchline. “ElevenLabs CEO opening X this morning to find his entire product released with open weights,” posted @shiri_shh — 6.5K likes for a joke that isn’t really a joke.

The open-weight wave is getting loud in the most literal sense. Vijay Shekhar Sharma quote-posted the claim of “the world’s best Indic voice AI model, fully open-weight,” flagging it for the India desk: high-quality voice in Indian languages, free to take. The pattern rhymes with everything Isenberg said about the model layer commoditising — voice is simply where it’s happening fastest, because the eval is trivially legible. You don’t need a benchmark table. You play the clip.

• Voice commoditises first because quality is audible in seconds — the fastest legible eval in AI, and the hardest to defend with marketing.

• Open weights are a go-to-market strategy now, not just a research gesture — they arrive as product launches with their own memes.

• Indic languages are the open frontier: fully open-weight Indic voice models lower the cost of building for the next 500M users to near zero.

FishAudio’s Drama 3 · The ElevenLabs open-weight joke · Open-weight Indic voice

How to Be More Ambitious: A Tutorial for Humans with Models

Felix Rieseberg, an engineering lead working on Claude, noticed that “be more ambitious” is the kind of advice everyone gives and nobody can follow — so he wrote the tutorial he wished existed. His essay breaks ambition into three trainable dimensions: Altitude (aim at outcomes, not tasks — crash-free apps instead of fixed crashes; redesigning the review instead of filling its template), Depth (expert-grade execution: the pros’ instruments not just their vocabulary, unreasonable goals, adversarially checking assumptions against raw data, an outside judge), and Range (cross-field thinking: assemble expert “crews,” swap nouns for verbs, never empty roleplay). All of it assumes the model has a computer and data access — ambition is now a tooling question.

The closing argument is the one to keep. Rieseberg reaches for the psychology of self-efficacy: belief in possibility drives accomplishment — Dantzig solving the “unsolvable” homework, the four-minute mile, The Inner Game of Tennis, Linux. As capabilities leap, he argues, we need deliberate techniques for bigger goals, because the bottleneck moved from what’s possible to what we dare to attempt. Scott Farkas’s gloss on the post is the TL;DR the feed deserved: most people aren’t using today’s AI to its full potential. 6.7K likes, 13.1K bookmarks — the most-bookmarked essay of your week, which suggests the bottleneck is real and widely felt.

• Altitude: restate the task one level up — from fixing crashes to shipping crash-free, from filling templates to redesigning the ritual.

• Depth: use the instruments, not just the vocabulary — unreasonable goals, adversarial checks against raw data, a judge outside the loop.

• Range: assemble expert crews and swap nouns for verbs — roleplay is cheap, cross-field structure is not.

• Self-efficacy is the meta-skill: Dantzig, the four-minute mile, Linux — belief in possibility keeps being the binding constraint, now with better tools.

Felix’s post · The full tutorial · Farkas’s gloss

Railway Hacked AI SEO: $20M to ~$100M ARR in One Year

Nakul Mandan’s line set the table: Railway started the year at $20M ARR and is now at roughly $100M — a solo founder, fully remote, about forty people. Ira Bodnar’s thirteen-point breakdown of how is this week’s most actionable thread, because the channel isn’t ads or outbound. It’s AI recommendations. Claude Code now picks Railway for 82% of Python apps, and the founder treats that as a growth channel the way a previous generation treated Google.

The playbook is thirteen moves, all legible: 6,653 how-to pages for self-hosting tools like n8n with one-click Railway setup. 127 framework deploy guides — Next.js, Django, Rails, FastAPI, Laravel, Go. Comparison and migration pages against Heroku, Vercel, Render, Replit, Lovable and five more. A help centre rewritten for AI: an index file plus plain-text versions. Agent-oriented docs for one-step Claude Code and Cursor setup. ChatGPT and Claude Code plugins for chat-driven deploys. And the killer detail: “ssh railway new” — a no-signup server for agents. No human onboarding, no friction; the agent provisions its own box. A weekly changelog since 2021 — 312 pages — keeps the corpus fresh, and every new model release lifts signups; there was no summer slowdown this year. All of it run by a seven-person GTM team.

• AI recommendation is a growth channel. 82% pick-rate inside Claude Code is distribution no ad budget buys — treat it like a ranking to win.

• Write docs for the reader that matters: plain-text, indexed, agent-oriented pages plus plugins beat pretty docs nobody’s model can parse.

• Remove the human from onboarding: “ssh railway new” — a no-signup server for agents — is the logical endpoint of product-led growth in the agentic era.

• Freshness compounds: 312 weekly changelogs since 2021 keep the corpus alive; each model release lifts signups like clockwork.

Bodnar’s 13-point thread · Nakul’s original post

Look at People’s Eyes: The Best PMF Definition Ever

Anuj Rathi’s verdict is the headline: “The best version of PMF definition ever!” Lenny Rachitsky’s version: when you’re showing people early prototypes to see if you’ve found it — “Look at people’s eyes. That’s the metric.” No survey, no NPS, no cohort chart. The involuntary widening of someone’s eyes when the prototype does the thing — that’s product-market fit before the spreadsheet can see it.

It lands well next to Lenny’s own post from the 11th, which Rathi’s thread echoes: every company needs to think and operate like a labs team now. The lab posture — experiments, eyes on users, fast prototypes — is no longer the privilege of research orgs; it’s the operating model for everyone, because the cost of a prototype just fell to near zero. And for the company you’re considering joining (or building), Lenny’s bookmarked share this week is a plug-and-play moat-check skill — an artifact that pressure-tests a company’s moat, with Ramp as the worked example. Eyes for PMF, a checklist for moats: the two cheapest diligence tools in the business.

• Eyes are the leading indicator. Prototype demos where users lean in beat every survey — watch the face, not the form.

• Operate like a lab. When prototypes are free, the differentiator is experimental velocity, not planning quality.

• Check the moat before the mission. Lenny’s moat-check skill turns diligence into a repeatable artifact — steal the format for your own evals.

Rathi on Lenny’s PMF line · Lenny’s labs-team post · The moat-check skill

Agents get observability: OpenTelemetry goes beta

ClaudeDevs put OpenTelemetry traces from the Messages API, Claude Managed Agents and the Claude SDKs into public beta — your agents’ traces land in your own observability stack. This is the harness era’s plumbing: you can’t govern what you can’t see, and failure logs are becoming the valuable asset Isenberg described. Ship the traces before you need them. ClaudeDevs’ announcement

Claude Code Desktop: subsessions feel like “v2”

Thariq’s read on Claude Code Desktop: it’s now aware of its own capabilities — spinning off subsessions, working with you — “almost like a v2.” The agent that manages its own parallel workers is the harness pattern reaching the desktop. Watch how subsessions change your own prompting: delegate, don’t dictate. Thariq’s post

Glass, via Delian: the third-party advantage

“Everyone thought Ramp was just corporate cards” — Delian Asparouhov’s verdict on Glass: a Codex/Cowork competitor that’s actually better because it’s a third party trying to save you money instead of a lab selling you tokens. Incentive alignment as a feature. Keith Rabois’s one-liner: “Glass empowers every company to operate on Ramp time.” Delian’s take · Rabois on Ramp time

Zuck picked Wang after talking to every lab leader

Shruti Mishra’s clip: Zuckerberg talked to every AI lab leader before picking Alexandr Wang — and Meta even tried to buy Ilya. The talent market at the frontier is a handful of names and one decisive buyer. Pair with the Muse Wonder saga from recent editions: the lab-leader circuit is where strategy actually gets set. The 2:35 video

Apple keeps ~$450 of every $1,000 iPhone

Vijay Shekhar Sharma, quoting Simon Kuestenmacher’s margin teardown: Apple keeps roughly $450 of every $1,000 iPhone — “While it is only $30 for India … but it is same as China.” The hardware margin map in one line: the value accrues to the platform owner, and India’s slice is the assembly slice. Remember this the next time someone pitches you “the iPhone of” anything. Vijay’s post

Apple forgot AI; the Gemma gang didn’t

“Apple forgot ai is the technology to work with — google/gemma gang perfectly aligns to the opportunity,” Vijay again, quote-posting DigUp: an open-source macOS app where Spotlight actually understands the meaning inside your files. Open weights plus local models are doing the sherlocking Apple won’t. The DigUp post

Harshil on ovens: the unserved market

Harshil Mathur’s memory: growing up, most Indians couldn’t bake at home — ovens were largely limited to commercial use. The point isn’t ovens; it’s how many “normal” markets are one appliance, one price point, one distribution unlock away from existing. The best opportunities look boring until they don’t. Harshil’s post

Zuck: “that’s just my face”

Mark Zuckerberg, replying to @Wahndo_, on whatever the screenshot showed: “that’s just my face.” 8.2M views for four words — the CEO-as-creator era in its purest form. Musk’s Argon-week line, proved in the same week by the man it was about. The reply

The heartbreak thread

Paras Chopra: “This hacker news thread is absolutely heart-breaking. Whoever thinks ‘humans will find a way’ ought [to] read it.” No further comment needed — some threads you read to calibrate your optimism against other people’s reality. Paras’s post

Play Store learns curation

Sameer Samat: Google Play is evolving to help people discover trusted, high-quality apps. When supply goes infinite, the platform’s job flips from hosting to filtering — the same curation thesis behind the Argon launch loop. Samat’s announcement

Argon’s video-gen clip, on repeat

Pichai quote-posting @fofrAI’s Argon video-generation clip — watched “a few times, especially to the end” — is the whole modern launch in miniature: the model ships, a creator finds the magic frame, the CEO amplifies it. Pichai’s quote-post

The moat-check you bookmarked

Your one bookmark-only save this week: Lenny’s plug-and-play moat-check skill — share the company you’re considering, get the moat interrogated, Ramp as the worked example. Keep it next to the eyes test. The moat-check skill