Don't marry a model

The Best Model Lasted Three Days

A week ago, no CTO had this on their risk register: the AI model your company runs on can be switched off by people who don't work at the lab that built it, and don't work for you either.

On June 9, Anthropic shipped Claude Fable 5, the most capable model it had ever made public, the kind that tops every rival on the standard tests and had Andrej Karpathy, one of AI's most-followed researchers, calling it a "major-version-bump-level step change" (developer-speak for a giant leap). Three days later it was gone. The U.S. government issued an export-control directive citing national security, barring access for any foreign national, including Anthropic's own staff. Since the company can't reliably check who's a foreign national and who isn't at the login screen, it pulled Fable 5 (and its locked-down sibling, Mythos 5) for everyone and rerouted to the older Opus 4.8. The trigger, as best anyone can tell, was a researcher getting Fable to find and patch security holes in code. Anthropic argued the same trick works on GPT-5.5 and that the order never spelled out the real worry, but none of it brought the model back. The best tool on the planet turned out to have a three-day shelf life.

The people caught flat-footed are the ones who spent two years hard-wiring everything to a single model. The CTO who standardized the whole company on one lab learned that the choice was never just a vendor decision, it was a geopolitical one, and the lab itself doesn't fully control the outcome. Pick your model, build everything on top of it, and you've handed a kill switch to a room full of people you'll never meet.

So the obvious move is to stop depending on any single one of them. OpenRouter's Fusion launched the same week Fable died: instead of sending your question to one model, it puts it to a panel of the best ones and has a judge model fuse their answers into a single response, which in OpenRouter's own testing beat any of them working alone. Cheaper tools do a lighter version, quietly swapping to whatever model is online and affordable when your first pick goes dark. The bigger bet, the one Box CEO Aaron Levie made out loud, is on open-weight models: one you download and run on your own machines is one nobody can repossess.

For three years the winning move was to find the best model and pour everything into it. Fable just proved any single model can vanish over a long weekend, which means the teams that come through the next scare are the ones that never tied themselves to one name to begin with.

Find Your Company on the Software Factory Ladder

Ryan Carson runs what he calls a "Code Factory," and it's basically a software company with one human in it. Agents write the code, a second set checks it, automated tests run on their own, errors get sorted, and something keeps an eye on the live product, while Carson keeps the one job he wanted: setting the rules the machines work under. He says he's shipped more than a thousand code changes this way, the output of a whole team from a single person.

This isn't a fringe stunt. Anthropic says more than 90% of its own code is now written by AI, Google says about 75% (up from 25% a year ago), and at OpenAI nearly every engineer works through an internal agent. Here's why it matters even if your company will never ship a line of code: engineering went first only because its work is easy to check. A test passes or it doesn't. The moment a kind of work has a clear right answer a machine can verify, it gets put on a line. Karpathy put the rule cleanly, that old software automates what you can specify and AI automates what you can verify. Finance reconciliation has a right answer. So does a resolved support ticket, a clean data feed, an email that either sent or didn't. Engineering is just the first function on the belt, and the one worth losing sleep over is which one in your shop is next.

The clearest way to find your spot, from a breakdown Alex Lieberman (Morning Brew) wrote for non-technical leaders, is a five-rung ladder. You can place your team in under a minute:

• Level 0, Artisan: every step is a human. AI is nowhere.
• Level 1, Assisted: engineers use AI to type faster, but a person still does and approves every step. Most companies are honestly here.
• Level 2, Delegated: an agent writes the change and submits it for approval on its own, but a human signs off on every one. Most "AI-forward" teams are stuck on this rung.
• Level 3, Supervised Factory: low-risk changes ship with no human in the loop, because a second agent reviews the first one's work and the automated tests have to pass. People set the guardrails and handle anything touching money or security.
• Level 4, Autonomous Factory: the bug gets caught, fixed, and shipped before anyone notices, and the humans are off deciding what the product should become. Almost nobody is really here.

You climb by pulling the human out of one more step, and only as fast as your safety net can catch a mistake. The four moves below are aimed at whoever runs your engineering, but skim them so you know what to ask for.

1. Hand one agent a whole ticket: If you're on rung 1, AI is still just fancy autocomplete for your team. Take the best-tested, lowest-risk corner of your code, give an agent an entire bug to fix end to end, and have a person review it. Do it with your next ten bugs.
2. Write down what "low-risk" means: Name the changes that never ship without a human, nothing touching payments, logins, or customer data, and let everything outside that fence move faster.
3. Stand up a reviewer agent and trust your tests: Put a second agent on reviewing the first one's work, require the automated tests to pass, and let anything that clears both ship on its own. This is the jump that needs real infrastructure: a gate that only lets a change through once the reviewer agent and the tests both sign off, plus a fast way to roll back when one slips. Per Google's DORA research (a long-running study of what makes software teams fast and reliable), a solid internal platform is the biggest thing separating teams that win with AI from teams that drown in it.
4. Know when to stop: Pulling the last human off the trigger is about trust and monitoring, not a better model, and most companies shouldn't sprint here. Anything touching money, safety, or someone's data deserves a person in the loop for a long while yet.

The percentage of your code written by AI is a vanity metric. The number that actually predicts who wins is how much of your line runs with nobody standing on it.

Also worth your time

OpenAI wants to mint 300,000 consultants

OpenAI launched a Partner Network with a $150 million budget and a goal to certify 300,000 consultants by year-end, starting with Accenture, Bain, BCG, McKinsey, and PwC. The same firms whose junior-heavy, bill-by-the-hour model AI is quietly eating are now lining up to resell it.

Salesforce bought its way into customer-service robots

Two weeks after Benioff said he'd stopped hiring engineers, Salesforce agreed to buy Fin, the customer-service AI company formerly called Intercom, for about $3.6 billion, and fold its agent (which resolves roughly 76% of tickets on its own) into Agentforce. The company that sells everyone else "build your own agents" just bought one instead.

Get paid to wait

The Claude Code progress spinner might be the most-watched loading bar on Earth, so Andrew McCalip turned it into an ad marketplace called Kickbacks: advertisers bid to run ads while your agent works, and you keep half. Half the internet called it genius, the other half begged him to get ads out of their terminal.

A reviewer with taste, on demand

Gokul Rajaram launched Coach, an opinionated AI reviewer for product docs, strategy memos, and roadmaps, built for the question AI made scarce: not whether you can ship it, but whether you should. When making things gets cheap, judgment is the whole game.

The next moat is the loop you own

A take making the rounds argues the real prize now is owning the feedback loop that turns your company's own processes and hard-won knowledge into a system that keeps getting better at your specific work. Everyone can rent the same raw intelligence now. The lasting edge goes to whoever teaches it the things only their company knows.

See you next Tuesday. Sven