Replit's 'Self-Driving Company': Agents Do the Legwork, People Make the Calls
Replit CEO and co-founder Amjad Masad says the future of work is the "self-driving company," where AI agents handle routine work and people are valued for judgment and taste. He spoke at TechCrunch Disrupt in San Francisco about how Replit has reorganized its own operations around that model. TechCrunch
Masad said Replit changed how it operates entirely. The closest analogy is a self-driving company where agents do much of the driving. It does not mean autonomy for its own sake. It means agents carry tickets, code context and operational state far enough forward that human time shifts to decisions only people can make.
Support is the example he described in most detail. Incoming support tickets are first examined by an AI agent that performs initial triage, meaning sorting and prioritizing, using GitHub issues, the codebase and Linear tickets. GitHub issues are past bug reports and discussions. The codebase is the actual program code. Linear tickets are the team task tracker. That grounds the response in the same materials engineering uses, with no separate knowledge base to drift and no copy-paste from stale runbooks.
When a support staffer decides to submit a patch, it is "just a click to deploy another agent." Triage, code search and fix generation run as one continuous loop. A human still authorizes the action. The agent does the search across issues, code and task tracking that used to take an afternoon.
Masad said Replit staff now spend more time on discussions, whiteboarding and architecture to decide what to work on. Roadmapping replaces ticket-chasing. Specification replaces scavenging for context.
How Replit defines the model
Replit laid out the idea outside the Disrupt stage. The company published a first-party post titled "The Self-Driving Company," and Masad expanded on the thesis in a separate write-up. Replit
The definition has two limits. Replit states a self-driving company is not one without people and that people still choose the destination. It separately describes a self-driving company as running on agents that act without a human re-checking every number. Replit
Taken together, the model is not lights-out automation with no staff present. It is delegated execution with human direction. According to Masad's description of the internal deployment, agents investigate production incidents, review pull requests, answer questions and analyze business data. Production incidents are problems in live software. Pull requests are proposed code changes. The human role moves to intent, prioritization and acceptance criteria, meaning what should be done, in what order, and what counts as done.
Masad described a CEO as a "glorified router" and said the self-driving company thesis is where the future lies for Replit. He linked that thesis to an enterprise ambition, saying Replit can create the operating system for the enterprise.
Enterprise context
Replit was founded in 2016 by Amjad Masad, Faris Masad and Haya Odeh and is based in San Francisco. The company describes itself as the world's leading online programming environment. It reached a $3 billion valuation on $150 million in annualized revenue last year. TechCrunch
Visa invested in Replit to power agentic payments for developers. Once agents leave the IDE, the coding tool, that work touches authorization, money movement and audit trails. TechCrunch
Replit uses agents as the link between systems of record, not only to draft code. GitHub issues hold history. The codebase holds ground truth. Linear holds ownership and state. The support queue handles intake. Pull requests and production telemetry, live signals from running software, handle verification. Each handoff that used to need human glue becomes an agent call with context attached.
The broader context here is familiar to anyone who lived through the shift from on-premises ticketing to chatops to platform engineering. Each generation lowered the cost of moving context between tools, and work reorganized around the new bottleneck.
In my view, Masad has named that bottleneck as judgment. When triage and first-draft fixes are cheap, the scarce resource is knowing what is worth fixing, what good looks like and what risk is acceptable.
Worth flagging for engineering leaders is the trust boundary Replit draws. Agents act without re-checking every number, but humans still choose the destination and still click to deploy the fixing agent. That keeps a person in the authorization path while removing the person from the search path. Full autonomy breaks auditability. Full manual review breaks throughput.
What this could mean for other teams is starting where context is already structured. Support to code to ticket is a tight loop with clear inputs and observable outputs. Incident investigation, pull request review and internal data analysis follow the same shape. Whiteboarding and architecture get the reclaimed hours, which is where senior engineers add the most leverage. If the bet holds, the long-term payoff is less about headcount and more about cycle time. Small teams can carry larger operational loads because agents carry the state between systems. That has, on balance, been the direction of good tooling for forty years, from compilers to version control to CI. Agents are the next layer for carrying routine work, leaving people to decide what matters.
