Why companies are becoming a series of loops | Anish Acharya (a16z)

There’s a lot of dread around AI: if you’re not mastering every new tool, do you get left behind for good? I wanted to test that “permanent underclass” meme and ask what it means for careers, companies, and ordinary life.

Not very much, man. Silicon Valley has this dark fantasy even as opportunity spreads and viable players multiply. Coding agents looked like they might become winner-take-all; instead, loads of products work. Jobs and the technical picture don’t support one player gaining an infinitesimal edge and never losing it.

That tracks. Every supposedly shocking milestone is visible, inspected, and followed by more iteration. It feels like a slow takeoff, not, “Tomorrow it becomes superintelligent and we’re all in trouble.”

Exactly. Fast takeoff usually means, “Everything proceeds as it has, then an unspecified thing happens.” Models improve absurdly fast, but benefits spread slowly through the economy. Physical systems, operations, and incentives constrain plenty of work. Give a pizza chain a data center full of PhDs; I’m not convinced it becomes exponentially better at pizza.

We under-credit employees too. We imagine Dilbert managers and low-agency consumers while assuming our own job is uniquely safe. People want leverage. At Kavak, mechanics go through Jedi Academy and, after six weeks, put serious agents into production. The shift is redesigning organizations around AI, like factories rebuilding themselves for electricity.

I’d be less stressed and more empowered. Nobody wants a tidier four-trillion-dollar company; they want a forty-trillion-dollar company. AI can turn years of roadmap into months. That is every PM’s fantasy. So ship stuff. Claire is my muse: she tries things, risks embarrassment, and looks like the best version of herself.

I love that. Your other big idea is AI-native companies as stacks of loops: input comes in, work gets done, impact goes out. What does that look like beyond engineering?

Agents repeatedly use tools, memory, and instructions; connect them into larger loops. A bug gets reproduced, fixed, reviewed, maybe approved by a human, released, and communicated back. Businesses can build versions in marketing, sales, support, legal, and product. Eventually those outputs should tell a CEO something strategic needs to change.

Humans stay essential. A growth loop can generate experiments, measure versions, ship the statistical winner, keep a holdout, and begin again. It climbs a hill efficiently, then hits a plateau. Somebody has to do out-of-distribution thinking, pick the next hill, and be right. You still have to point Claude somewhere worthwhile.

When an agent fails, ask: what do I know that it doesn’t? At Kavak, an agent can call a person when stuck. The person coaches it, the trace is captured, and ideally that escalation disappears next time. If loops take downstream work off your plate, you get to hike, dream, and find the next helicopter climb.

Use cheap narrow models and expensive frontier intelligence where they fit. The question is whether a role’s upside is bounded. You can only close the books correctly. But a support call might reveal a company-changing problem. And don’t obsess only over job loss: AI separates desire from skill. Make music, software, a better family life—things that add texture, connection, and agency.

For founders, assume intelligence gets astonishingly cheap and rebuild the business from there. Don’t overthink a moat on day one; durability emerges through momentum, craft, engagement, data, brand, scale, networks, and a beloved product. My update is that ideas can be too small. Build the strongest form of the vision—the moon-sized outcome, or the moon-sized crater.

For product people, just make more things. Once a week is plenty, even if it’s tiny or for three friends. Build for learning and joy, not only an audience. Make something for somebody else. Don’t be despondent—build, tag me, show people. Use the models, follow the weird idea, and tell me what you built.