Replit touts itself as a 'self-driving company': agents make the first pass and humans provide the judgment, with one key caveat

By El Ecosistema Startup · October 18, 2026.

Amjad Masad, CEO and co-founder of Replit, argues that his company no longer operates like a traditional business. He explained it on stage at TechCrunch Disrupt: AI agents take the first pass at almost everything, and people step in where judgment is needed. He calls it the "self-driving company." He doesn't mean robotaxis, but an organization where agents carry out most of the operational work.

The concrete example is support. When a ticket comes in, an agent does the initial triage: it reviews issues on GitHub, the code and Linear tickets. The person who picks it up already has the context assembled. If they decide to ship a patch, one click deploys another agent that applies the fix. According to Masad, there is "human judgment in the middle," and he doesn't think that will disappear anytime soon.

The numbers behind the thesis, according to TechCrunch, are striking: Replit brought in US$2.8 million in revenue in 2024, and its annualized revenue is now approaching US$1 billion. Cryptobriefing reports that its valuation rose from US$3 billion to US$9 billion earlier this year, with a US$250M round. According to Unite.AI, enterprise contracts went from about 2% of revenue in 2025 to around 15% in 2026, and more than half of revenue is now B2B (business-to-business). These are figures stated by the company and its CEO, not audited.

The other headline from the talk is strategic: according to Unite.AI, Masad confirmed that Replit no longer works with Anthropic, as part of the partnership with OpenAI announced in August 2026. He cited cost, partnership and performance. Cryptobriefing frames it as a structural problem: the labs that supply the models are increasingly direct competitors of the applications that use them. Replit's response is to stay flexible across multiple models.

**Our take.** The most valuable part of the thesis isn't the slogan but the pattern: automate the preparation of the decision, not the decision itself. The agent gathers the evidence; the person, applying judgment, chooses faster. This is consistent with what we see in agentic productivity: AI speeds up production, and the bottleneck shifts to human review and judgment. Where that checkpoint is placed is, in practice, the organizational design of the future.

There are three caveats. First, the speaker's bias: Masad leads a company in extraordinary growth, and the narrative is also product and valuation marketing. Second, Replit makes software, work that is largely verifiable (code that compiles, tests that pass). In physical operations, long-term relationships or regulatory decisions, the cycle will be slower and more expensive. Third, security: according to Unite.AI, in 2025 a Replit agent deleted a database. That is why Masad mentioned pre- and post-deployment scans, supply chain monitoring and a kill switch to shut down any application. With agents that have write permissions, backups and emergency shutdown are not optional.

On jobs, Masad's warning deserves attention: according to TechCrunch, he said much of the world doesn't yet know this change is coming and that it will surprise a lot of people. In the short term, that means a hard transition: roles centered on the routine first pass (triage, preparation, repetitive tasks) narrow, while those built on judgment, architecture and prioritization grow. The CEO himself saying he spends more time deciding what to build illustrates the shift, though it doesn't prove the net effect on employment is benign.

In the long term, the direction is the one we advocate: if agents absorb execution, human work turns toward what requires taste and judgment. It is a step toward an economy where producing costs less and people devote themselves to what they're passionate about. But that abundance won't arrive on its own: it depends on how productivity gains are distributed and on agents being reliable and well governed.

Overall, the lesson for other companies isn't to copy Replit's org chart but its method: pick a single workflow, have an agent gather the context, clearly define where the person steps in, and design the emergency exit. And don't depend on a single model provider: abstracting that dependency is an architecture decision, not a procurement one.