Replit's 'Self-Driving Company' Still Keeps Humans in Control

Replit CEO Amjad Masad calls his company's new operating model a "self-driving company," but the example he gives is not a business running without people. It is a workflow in which AI agents gather evidence and perform routine work while a person makes the consequential decision.

When a support ticket arrives, a Replit agent can inspect GitHub issues, the codebase and Linear tickets before the request reaches an employee. If the employee decides a code change is needed, another agent can prepare the patch. Replit says it uses agents for production incidents, pull-request reviews, business-data analysis, support triage and sales research, with access to systems including GitHub, cloud infrastructure, Slack, Notion and Zendesk.

That access is what makes the model useful and risky. An agent that can read operational systems and propose fixes may shorten investigation time, but permissions, audit trails, testing and human approval become part of the product rather than back-office controls. Replit's support example keeps a person between diagnosis and deployment, an important distinction from full autonomy.

The company has financial incentives to prove this approach scales. Replit said in March that it raised $400 million at a $9 billion valuation and expected to reach $1 billion in run-rate revenue by the end of 2026. It also said users at 85% of the Fortune 500 were building with its tools. Those are company claims from a private business, not audited public-company results, but they show why Replit is pushing beyond individual coding into enterprise workflows.

Masad argues that agents shift employee time toward architecture, prioritization and judgment rather than eliminating people. The evidence offered so far is operational and anecdotal, not a controlled productivity study. Replit has not disclosed in the cited material how support resolution time, defect rates, headcount or gross margins changed after adopting the model.

For technology investors, the broader question is whether agent platforms can turn access to many business systems into reliable, governed work. Replit's own deployment suggests demand for orchestration, identity controls, observability and security alongside the models themselves. The strongest proof would be measurable customer outcomes and a record of safe execution, not the "self-driving" label.