Anish Acharya argues company building is becoming a series of agent loops, with a coding bug-fix loop capable of running from report to shipped fix in five minutes.
Acharya says Sundar Pichai wants a $40tn Google rather than a more efficient $4tn one, and a Google executive reported roadmap work accelerating rather than layoffs.
Acharya argues frontier models are irrationally priced, since Fable 5 costs 100 times more per IQ point than Opus 4.8, so the market will split between cheap open-weight models and premium frontier ones.
Companies will run on agent loops, with humans supplying ideas
Acharya argues AI is spreading as a broad, slow economic shift rather than a winner-take-all race or a sudden takeoff. Companies will become series of agent loops that finish work in minutes, while people supply new ideas and handle exceptions. That favors ambition over headcount cuts, because the gain comes from shipping more and better ideas.
Takeaways
AI is not heading to winner-take-all. Each part of the stack now has around 20 relevant players rather than two. Claude Code, Codex, Lovable, Replit and Wabi are all working, though two years ago coding agents looked winner-take-all. Progress looks auto-catalytic, not recursively self-improving.
Companies become series of agent loops. An agent is a model running in a loop with tools, memory and skill files. A bug report to shipped fix can take five minutes. Loops are expected to reach marketing, sales, support and legal.
Loops plateau, so humans supply new ideas. Loops climb to local maxima and then stall, so people provide out-of-distribution thinking and handle exceptions. At Kavak, agents call a human when stuck and learn from the coaching. A growth loop would generate, measure and ship every variant.
Economic diffusion is slow, so no sudden takeoff. Acharya doubts a fast takeoff because many problems are less intelligence bound than assumed. Electricity took 40 years to reshape factories. Kavak, the used-car seller in Mexico, trains all staff, including mechanics, to build agents. Job postings are higher than ever.
Ambition, not headcount cuts, is the response. Sundar Pichai reportedly wants a $40tn Google rather than a more efficient $4tn one. A Google executive said no one was laid off and two years of roadmap now happens in three months. Simulated ideas let the best one win.
Frontier models are irrationally priced. One extra IQ point from Fable 5 costs 100 times more than from Opus 4.8. Acharya expects a split between open-weight mid-IQ models for narrow jobs and frontier models for sales, support, research and engineering.
Consumer AI is a product design problem. More people want to spend time than save time, so the challenge is design, not models. Acharya says consumer AI is at an iPhone 2010 stage. Startups can build disagreeable products that incumbents avoid.
Moats and staying current
Moats are discovered, not designed, a line Acharya borrows from Jesse of Decagon. Cursor moved from a high-NPS product to training its own models.
Classic moats such as network effects still hold, and they have returned to grassroots word of mouth. Rachitsky adds that distribution is a new moat.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
a16z would back a product with no clear moat story if it shows momentum, craft and growing engagement.
Product people should pick a side project, use each new model to build something, and ship it and talk about it.