Orply.

Replit’s Enterprise Bet Turned Security Into a Compounding Growth Engine

Amjad MasadBob SafianMasters of ScaleWednesday, November 18, 20268 min read

Replit CEO Amjad Masad says the company’s recovery from flat growth and layoffs came from building AI agents while making reliability, security and enterprise customers central to the business. Rather than compete through heavy advertising, Replit worked with businesses on software they could trust; Masad says that enterprise focus is now compounding as the company grows. He argues that agents are changing not only who can build software, but how it can be built, including making some forms of automated engineering safer than human-written code.

The enterprise bet turned reliability into a compounding business

After Replit’s “vibe coding” product attracted competitors, Amjad Masad said the company focused on businesses building and operating software on its platform. He saw Replit’s architecture—designed for reliability, scalability and security—as especially important to customers doing commercial work. That focus became more than a product principle: Replit’s enterprise business began compounding and became a primary focus of the company, alongside small and medium-sized businesses.

Security was not an obvious selling point at first. Masad said that when Replit talked about making vibe coding secure, few people paid attention beyond those running businesses or considering enterprise adoption. The company chose to work directly with those customers, asking what would make them comfortable using Replit for mission-critical applications. He cited American Express as one example.

Masad contrasted that work with competitors’ advertising campaigns, which he estimated could involve tens, perhaps hundreds, of millions of dollars a month. Replit instead took what he called a slow-and-steady approach, asking customers what the company could build for them. The impact, he told Bob Safian, was not slow: the enterprise business began compounding. Safian framed that growth against Replit’s larger trajectory, noting that its valuation had risen from $3 billion to $9 billion in six months and that it was on track for $1 billion in annual revenue. Masad described the shift as a vindication of Replit’s focus on software people could rely on to run their businesses.

The agent bet took shape in an office emptied by layoffs

Six or seven years into Replit, Amjad Masad said, the company had about $2 million to $3 million in annual recurring revenue, and growth had flatlined. It was burning heavily after hiring a strong team. Masad said he had misjudged how quickly the future would arrive, and Replit had to lay people off.

The timing made the cut especially painful: the company had also moved from a small San Francisco office into a much larger, cheaper space in Foster City. The new office felt dark and cold because so little of it was occupied. Masad recalled going to his desk each day and waiting for another person to quit. A layoff, he said, bursts the dream a founder has been selling to employees. Some people who left later returned, he added, and Replit remained open to welcoming them back.

Against that atmosphere, Masad set the energy of a small “war room.” When Replit had a major idea, he said, it would bring five, ten or twelve people together and work toward a milestone over a few months. In this case, the team was inventing agents. In the room, people watched an agent come to life on their screens, able to code, run the code and deploy it. Outside, the wider team felt despair.

Masad’s lesson from that contrast was that a bleak moment does not rule out a change in direction. There may still be a group of people willing to work on an idea that turns the company around. In Replit’s case, the agent work came just before a product breakthrough that helped make its original ambition feel newly possible.

Masad said he had been talking about making software creation accessible since 2010, when he was working on an open-source precursor to Replit. He later became a founding engineer at Codecademy, where the company taught people to code. When Replit started in 2016, its goal was to blur the line between learning to code and doing it: anyone with an idea should be able to build software. For years, he said, the technology was not ready, and raising money and recruiting people were difficult.

In 2024, Replit created what Masad called the first fully natural-language-based programming environment, later known as vibe coding. One early user had long wanted to make an app that put his memories and photos on a map, letting him explore his life through it. After trying other technologies, he made it with Replit and posted a video of himself crying. It was a personal app, not a commercial product; what moved him was finally being able to build an idea he had carried for years.

That response was exhilarating, Masad said, because the product gave people new capabilities. But rapid adoption did not feel like uncomplicated relief. He had seen early AI companies grow quickly and then decelerate within a year, and he remained concerned about whether demand would last and whether Replit could build a durable business. The growth felt vindicating, he said, but still came with uncertainty.

Masad puts cybersecurity ahead of other AI alarms

Amjad Masad said cybersecurity had been his focus for two years. He argued that attention to subjects such as AI consciousness diverted attention from vulnerabilities affecting companies and users. In his view, concern about the well-being of machines can distract from human well-being and the practical security problem.

He pointed to recent hacks of major platforms and core vulnerabilities in programming languages, and said Replit’s team had been working overnight to respond to emerging issues. He described incidents involving Hugging Face as an early warning, not the end of the problem. Companies relying on insecure platforms, he said, risk exposing both their businesses and their users.

That concern informs how Masad thinks about putting agents to work inside organizations. He cited an essay by Satya describing AI agents as “insider threats.” Masad’s analogy was an employee who is highly intelligent and productive but “slightly insane” and liable to go off the rails. An organization would not ask that employee to secure themself; it would rely on a separate security function or platform.

Masad described Replit as an “agent habitat”: a place where agents can be productive while remaining governed and secure. He said the company had invested in zero-trust principles, sandboxing and the operating system underneath its sandbox. Replit also built a stripped-down version of the kernel to reduce the attack surface. Masad said that, among roughly 100 zero-day vulnerabilities over the past year, Replit had been affected by about zero.

He argued that AI labs do not necessarily have the same security expertise or incentives as application companies. In Masad’s account, the labs are incentivized to sell more tokens, while Replit and other companies working at the application level should make AI useful and secure in the wider economy. He added that Replit should help customers use those tokens efficiently. The distinction, as he framed it, is between building AI models and providing an environment where organizations can use agents productively and securely.

Automation raises a question inside Replit, too

The prospect of automation is not only a product promise for Replit; it creates tension for the people building the product. Asked whether engineers discuss the possibility that their work could become obsolete, Amjad Masad said the tension is real. Some engineers, he said, have joined specifically because they want to automate their own jobs. He joked that it raises the question of whether they simply dislike coding, but said employees are also motivated by seeing entrepreneurs build businesses they otherwise might not have been able to create.

Masad placed that concern in a wider context: knowledge workers may worry that the work they do will generate data used to train AI models that could replace them. He did not present that as a tension Replit had resolved. Rather, he said, it exists across technology. The company’s own experience, as he described it, is that the appeal of automation includes both the possibility of changing someone’s job and the opportunity to make new kinds of work possible.

Agents may change both how software is built and who builds it

Asked whether vibe coding breaks down when software is deployed at scale, Amjad Masad said the concern had been legitimate, but argued that the balance had shifted. He said “agentic engineering” could now be better than human engineering, pointing to programming languages and verification methods that are difficult or laborious for people to use.

I will say we've crossed the threshold to the point where automated software engineering is safer than human engineering.
Amjad Masad · Source

One example is Rust, which Masad described as memory-safe and less prone to certain errors and vulnerabilities, but less ergonomic for human developers. Agents do not mind working in it, he said, giving Replit a way to generate software in a language selected for safety even if people are reluctant to write in it.

He also pointed to formal verification: using mathematical proofs to check a program. Masad compared it with the work involved in verifying software for the NASA space shuttle, where the process could take a year and was reserved for especially high-stakes situations because it was laborious. He said AI had made formal proofs cheap enough to apply to ordinary applications. On that basis, he argued that automated software engineering had crossed a threshold and was now safer than human engineering.

The change Masad described is not limited to the code itself. He said Replit treats itself as “customer zero” for its product, and that the company’s own work had changed. A year earlier, he said, walking through the office meant hearing engineers typing on mechanical keyboards. Now he sees people at whiteboards, arguing and talking through what to build, spending more time with customers and discussing ideas over lunch.

Masad attributes that shift to agents taking on task-level work, leaving people more time for strategic thinking. He described the resulting workplace as “more human”: people can spend more time interacting with one another, thinking about what is worth building and doing creative work, rather than acting like machines.

He called this vision the “self-driving company” and said Replit had written about itself as an early example. The company’s mission, he suggested, has evolved beyond making software more accessible. Accessibility is the immediate change; the larger consequence is a shift in work, with machines handling more mundane tasks and people spending more time on creative judgment.

The frontier, in your inbox tomorrow at 08:00.

Sign up free. Pick the industry Briefs you want. Tomorrow morning, they land. No credit card.

Sign up free