Replit’s Early Infrastructure Let Its Coding Agent Work Until the Software Ran

Replit’s early advantage was not simply that it could generate code. Its agent could keep working through errors until a program worked, then provision infrastructure to run it. Amjad Masad says the idea grew out of watching beginners shuttle code and error messages between ChatGPT and Replit: they would paste in generated code, copy an error back to ChatGPT, and paste the suggested fix into the project. The opportunity was to automate that whole process.

That meant making the model interpret errors and continue acting rather than stop after one response. Paul Graham called this “go to completion.” Masad says Replit’s 2024 launch demonstrated that models could be run this way and could also provision infrastructure: the agent could call tools to start a database, spin up Redis, or host an application. Some model companies later contacted Replit, Masad says, surprised that the agent worked and asking how the company had built it.

The early system took more engineering than today’s versions because the models were less capable. Masad estimates the 2024 agent “harness”—the surrounding engineering that lets the model act—was ten times more complex than the harness needed now. As models improve, he says, less engineering is required to make them function as agents. That makes it easier for competitors to catch up, but Masad treats commoditization as a standing condition: a company has to accept that its current work will be copied or made cheaper and keep building what comes next.

Replit’s agent was unreliable at launch. Masad recalls that it failed perhaps 30% of the time. Users nevertheless tolerated the failures because the experience felt magical and because their alternative—manually moving code and errors between tools—was also frustrating. For people who could not program at all, a system that worked most of the time could still be a substantial improvement.

Graham says the agent unsettled him when Masad demonstrated it in the summer of 2024. He had spent decades treating programming as a craft of carefully written code; Masad’s instruction was, “Don’t look at the code.” Graham describes realizing that the code he had spent years making was “just object code.” Masad says he had already grieved the loss of the craft by 2020. He expects engineers to pass through denial and depression before acceptance.

Amjad Masad describes Replit’s shift to agents as an extension of what the company had already built, even though it looked abrupt to outsiders. The company had spent years making programming work online and collaboratively, including the underlying systems that let users run and deploy projects. When Replit put an agent into a project, the agent could be treated as another user in the same environment. It could work alongside a person because the collaborative system was already there.

The organization followed the same logic. Replit had teams for code execution and sandboxing, deployment and cloud services, the interface, and the connections between them. The company added an agent team in the middle. Infrastructure teams began exposing tools for agents; the IDE team began building ways for agents to edit code.

That editing work was especially difficult in 2024. Models hallucinated line numbers, so Replit could not reliably ask them to replace code at a particular line. They could also leave unfinished placeholders—Masad gives the example of a function followed by “to do, finish it later”—or stop too early. Replit had to detect such behavior and prompt the models to continue. Paul Graham notes that early agents needed robust editing tools because the models were still poor at the work; Masad says agents can now operate through a shell and ordinary commands.

Masad says Replit employs kernel and operating-system engineers and built its own file system to make changes transactional and reversible. He connects some of this work to the company’s earlier effort to make coding collaborative, which required inventing core infrastructure before agents could use it. The systems work can be overlooked in a simple-looking product, but it matters when a company has to control what its agents can do.

That depth matters when the customer is a bank or manufacturer rather than an individual experimenting with an app. Masad says enterprises are more sensitive to reliability and security, and that competitors sometimes build applications in reckless ways. Replit adds guardrails that can make the product slower or more expensive, he says, but enterprise customers value them.

Masad says Replit expected AI autocomplete to write a large share of the code even before full agentic coding—perhaps 50% or 60%. But the company’s longer horizon preceded both autocomplete and the Transformer era. Its 2015–16 pitch deck proposed that a platform with millions of people coding could generate enough activity to train a model that predicted code.

The aspiration came earlier still. Masad traces it to 2008, when he wished he could program on his Symbian smartphone while sitting in class. Replit later built a mobile programming app, an idea Graham recalls people laughing at. For Masad, the thread was a durable goal: make programming available in more places and to more people.

Graham recalls telling Masad that making programming easier could increase the number of programmers by far more than the percentage of the improvement. Masad says Replit eventually had to go “the whole way” to make the company’s mission real. Graham argues that Masad’s predictions had been sharpened by thinking about the problem for nearly two decades.

Amjad Masad says some of Replit’s first paying customers for the agent had previously hired outsourced developers to build an app for around $100,000; with Replit, they could get something done for about $1,000. He sees this as evidence of demand that programmers might dismiss if they judge the resulting software only by its quality. Paul Graham admits he had seen outsourcing shops as proof that people were getting bad software, rather than as proof that customers had a problem they were willing to pay to solve.

Replit had earlier tried to serve that demand through Replit Bounties, a marketplace in which people could pay others to complete programming tasks. Masad says the people doing the work were already using Replit’s rudimentary AI, which gave them an advantage. But he shut the program down after concluding that agents would eventually let customers request the whole product directly, without an intermediary.

Masad says that shift is now playing out inside large companies. In banks, manufacturers, and other businesses, employees who have adopted Replit are replacing internal tools and legacy software. He reports that customers have described projects previously scoped at $10 million and completed for $250,000 using Replit. He says the largest impact is not necessarily in technology companies, which can build software themselves, but in ordinary companies that otherwise struggle to make or change the tools they need.

The people driving adoption are often ambitious employees inside those organizations: for example, someone in sales building a tool for a sales department. Masad frames this as consistent with Replit’s original mission, not a departure from it. Graham notes that people inside companies can be scrappy even when the companies themselves are not. Masad says employees who bring Replit into their organizations are often promoted because their work has visible impact.

Enterprise adoption brings its own constraints. Administrators and IT teams need controls over what employees can do and how much they can spend. Masad calls this a “principal actor problem”: the people using software are not always the people who decide whether to buy it. That gap can lead security and IT teams to make a product worse for its users.

One example is token budgets. Masad disagrees with giving every employee the same fixed monthly allowance, such as $100. Based on what Replit sees in its customers, he says, the returns from AI use are distributed unevenly: one person may produce 100 times the value of another, and the gap may grow as agents become more capable. Companies still need to control spending, but Masad argues they should be helped to see where valuable work is happening, rather than simply rationing use equally. For now, Replit often has to provide the controls customers ask for in order to get into an organization, then develop better ways to manage the trade-offs.

Amjad Masad says Replit no longer describes itself primarily as a vibe-coding company. Its broader idea is the “self-driving company”: a business in which AI agents increasingly create and operate the software the organization needs. Replit is trying to work through that change internally as “customer zero”—a place to learn how teams identify valuable work and use agents to do it.

The ambition follows from a change Masad says he did not initially appreciate. Making software creation more accessible does not only affect programmers. It can change how people across a company work, because more of the work can be done by creating software. Paul Graham observes that programming once seemed peripheral to the way most companies operate. Masad’s point is that software creation may become part of ordinary business work.

Masad says Replit should let customers show where the change is happening and respond to those needs, rather than announcing a sweeping strategy and expecting customers to follow. Graham describes the process as recognizing a pattern in one customer’s work, helping that person, and then making the capability available more broadly. Replit’s underlying systems give it a way into companies; observing what employees are trying to do can reveal what to build next.

Masad says Replit’s earlier “idea to app fast” positioning has been copied by competitors. Graham’s analogy is that they are reproducing an old frame of Replit while the company has already moved on. Masad agrees that the company has to keep adapting as AI commoditizes existing work.

The same pattern separates an insider’s gradual transition from an outsider’s sudden shock. Replit experienced its agent as incremental because the company had already spent years on online programming, collaboration, infrastructure, and AI assistance. To Graham, the demonstration looked like a discontinuity. Masad compares that difference to the history of Apple: a product can appear obvious after launch even though its creation depended on years of preparation and technology that was not previously ready.

Amjad Masad says his approach to the future is to make predictions actively, revisit them, and treat change as the default. He points to the continued growth of computing power as a useful frame. After reading Ray Kurzweil, he began thinking in terms of a log-linear chart of computing capacity. He recalls a formulation he saw elsewhere: “It’s not that AI brought on the compute, the compute brought on AI.” In his account, deep learning would not have happened without the computing resources becoming available.

That frame helped Masad look at GPT-2 in 2019 and imagine later forms such as chat, agents, and personal assistants. He says he makes predictions, revisits them, and tries to resist the instinct to focus only on reasons a new capability might not work. Paul Graham puts the distinction as a difference in default assumptions: many people treat the world as static, while Masad expects exponential change and builds toward it.

Masad connects that posture to a willingness to abandon a craft when computers become good at it. He says he has started over and rewritten things repeatedly rather than treating a particular skill or identity as permanent. For people facing AI, he says, flexibility and lifelong learning matter.

Masad says Replit sees many companies still far behind Silicon Valley’s use of agents. In a hackathon with a large telecommunications company, he recalls, a survey found that 80% of participants were anxious about AI beforehand. After an hour using Replit, the proportions reversed: 80% were optimistic and 20% remained anxious. He takes the result as a sign that experience can change how people feel about the technology, while warning that many companies have not yet experienced the tools now familiar in Silicon Valley.

He says many banks are still using versions of ChatGPT from around 2023, with little connection to company systems. Meanwhile, companies are starting to reconsider old software they want to replace. Graham asks whether even COBOL systems are being rewritten; Masad says the process has begun, though Replit had earlier declined requests to rewrite customers’ code for them. Its aim was to teach organizations to use the tools, while other AI companies now offer legacy-code rewrites directly.