What If Your Company Drove Itself? How Replit Uses AI at Work

Replit CEO and co-founder Amjad Masad says the future of work is the "self-driving company." AI helpers do much of the routine work, while people are valued for judgment and taste. He spoke at TechCrunch Disrupt in San Francisco about how Replit changed its own way of working. TechCrunch
Masad said Replit changed how it operates entirely. Like a self-driving car that handles the steering while the driver picks the destination, agents move work tickets, code details and current status forward. It does not mean automation for its own sake. It means people spend time on decisions only people can make.
Support is the example he described in most detail. When a customer asks for help, an AI helper first reads the request and looks at past problem reports on GitHub, the actual program code, and the to-do list in Linear. The answer is based on the same information engineers use. There is no separate help guide that can go out of date, and no copy-paste from old instructions.
If a support worker decides a code fix is needed, it is "just a click to deploy another agent." Sorting the request, finding the right code and writing a fix run as one loop. A person still approves the action. The agent does the searching across reports, code and task lists that used to take all afternoon.
Masad said staff now spend more time talking, sketching ideas on whiteboards and planning how systems should be built. They spend less time chasing tickets and hunting for background.
How Replit defines the model
Replit explained the idea outside the Disrupt stage. The company published a post titled "The Self-Driving Company," and Masad wrote more about the idea in a separate piece. Replit
The definition has two rules. A self-driving company still has people, and people still choose the destination. At the same time, it runs on agents that act without a person double-checking every number. Replit
So it is not an office with no people. It is helpers doing assigned work under human direction. Masad said the agents look into problems in live software, check suggested code changes, answer questions and study business numbers. People decide what to do, what matters most and what counts as done.
Masad called a CEO a "glorified router" and said the self-driving company idea is where the future lies for Replit. He linked it to a bigger goal, saying Replit can create the operating system for the enterprise, the main platform other business work runs on.
Enterprise context
Replit was started in 2016 by Amjad Masad, Faris Masad and Haya Odeh and is based in San Francisco. The company calls itself the world's leading online programming environment. It reached a $3 billion value on $150 million in yearly revenue last year. TechCrunch
Visa invested in Replit to power agentic payments for developers. Once agents leave the coding tool, that work touches permission checks, moving money and keeping records. TechCrunch
Replit does not only use agents to write code. It uses them to connect its work systems. GitHub holds past history. The code holds the truth. Linear holds who owns what and the current status. The support inbox takes in requests. Suggested code changes and live signals from running software are used for checking. Each handoff that used to need a person to copy information over can now be an agent call with the background attached.
The broader context here is a pattern seen before. Past shifts moved work from old ticket systems to team chat tools to newer work platforms. Each time, it got cheaper to move information between tools.
In my view, Masad has named the new bottleneck as judgment. When sorting and first drafts are cheap, the hard part is knowing what is worth fixing, what good looks like and what risk is okay.
Worth flagging for leaders is the trust line Replit draws. Agents act without checking every number, but people still pick the destination and click to start the fixing agent. A person stays in the approval step but not in the search step.
What this could mean for other teams is to start where information is already organized. Help request to code to to-do list is a short loop with clear inputs and clear results. Checking live problems, reviewing code and studying internal numbers work the same way. If the bet holds, the payoff is less about fewer workers and more about faster work. Small teams can handle more because agents carry the background between systems. That has, on balance, been the path of good tools for forty years, and it leaves people to decide what matters.


