Korva is a scroll of the internet’s best ideas.
Agents broke, so they built software
Tools that kept breaking
They started on OpenClaw, where the agents started breaking and hallucinating, which got too frustrating. Claude Cowork came next and worked, though it felt limited. Hermes Agent and Perplexity Computer were the other two tools people bounced between. The firm had told everyone to token max it, with a couple thousand dollars a month on the table, because the gains were what mattered.
Building an intranet
Lovable came after that, and people used it to build intranets for the venture business. No venture business would spend half a million dollars and 12 months on an intranet; the usual move is to put it in Notion, Google Sheets, or something like it.
Software they now write
Then they kept building more and more software. Two or three people there now build software that five years ago would have cost maybe two or three million dollars a year with an outsourced company. Back then the firm wouldn't have done it, and a SaaS app shoehorned and customized to fit wouldn't have worked.
Why can AI now build internal software cheaply?
Large models trained on huge amounts of code can turn a plain-language description into a working app. Internal tools suit that well, since they have few users, forgiving requirements, and no public launch to get wrong. Researcher Andrej Karpathy coined the term vibe coding in a February 2025 post to describe building software this way.
Why did firms use Notion rather than custom software?
Custom software was costly to build and then costly to keep running. Author Robert Glass wrote in Facts and Fallacies of Software Engineering that maintenance consumes 40 to 80 percent of the total cost of software over its life. A flexible off-the-shelf tool let a team skip both bills and accept a rougher fit.
Why do AI agents break and hallucinate?
A language model writes each step conditioned on what it wrote before, so one wrong detail becomes part of the context for everything after it. Over a long task those small errors compound. OpenAI researchers argued in a 2025 paper titled Why Language Models Hallucinate that training and scoring methods reward guessing over admitting uncertainty.
More from All-In with Chamath, Jason, Sacks & Friedberg
