Replit explores the self-driving company model with AI agents

Masad's framing assumes agents will handle 'routine execution' cleanly, but the archive shows this assumption is premature. The real friction isn't structural redesign capacity; it's that current agents still fail at reliability, domain judgment, and ethical constraint when given autonomy.

This connects directly to Meta's memory coach work from early August, which revealed that even supervised agent architectures struggle with long-horizon reasoning and error correction. More broadly, the last week's coverage (OpenAI's research software agents, the Hugging Face breach, Greg Brockman's observation about employee resistance to AI intermediaries) all point to the same gap: agents can execute, but they can't yet be trusted unsupervised. Masad's vision of 'orchestration' sidesteps the harder problem that Replit would face: agents making wrong decisions confidently, or employees rejecting agent-mediated workflows entirely. The organizational redesign he describes assumes the technical problem is solved; the evidence suggests it's still the bottleneck.

If Replit ships agent-driven decision-making in production (not just internal tooling) within six months and publishes error rates and user adoption metrics, that's a real signal. If the company quietly keeps humans in the loop for high-stakes decisions or sees employee friction similar to what Brockman described, the 'self-running company' thesis collapses into standard automation theater.