How Replit's Amjad Masad Saw Vibe Coding Coming Years Before It Had a Name

The conversation opens with Masad describing how he expects people to react to AI. He said that with AI, everyone will have to go through a period of denial, then depression, and eventually acceptance. He then used Replit as an analogy: it is like an iceberg, with a visible part above the water where you tell it in English what to do and it writes your code, and underneath, a large amount of infrastructure for deployment and sandboxing.

The host then suggested that Replit describes itself as a vibe coding company. Masad responded that the company does not describe itself that way. He said Replit now sees itself as the platform for a transformation toward self-driving companies.

The host noted that earlier tools, including ChatGPT, generally worked in a single turn, and asked whether the system could loop indefinitely until the program worked. Masad confirmed that going to completion was the new capability. He said Replit's first demonstration showed that LLMs could be run to completion, which he believed was news to the model companies, who reached out to ask how it worked.

Masad also said the harness required to make the models act as agents was about ten times more complex in 2024 than it is today, because the models were less capable then. As models improve, he said, the engineering needed to make them into agents shrinks. He added that the company also showed AI could provision infrastructure, such as spinning up a database or a Redis host. That was different from competitors, he said, because Replit had already built the underlying infrastructure the agent could call.

Masad said that in Replit's original 2015–16 deck, the company discussed how a platform with millions of people coding could be used to train a machine learning model that predicts. The host said this was far ahead of its time, and noted it came before the Transformers paper. Masad connected this forecasting to his own experience as a child, when he saw his first Nokia phone and believed phones would become the new computer.

He also said Replit was the first programming platform with a mobile app, and that many people laughed at the idea of programming on a phone. He recalled that he used mobile access to fix bugs in the Hacker News list, though he said this was not voluntary and was simply the only way he could get internet access.

Asked when he first began thinking about building something like Replit, Masad said it started when he had a Symbian smartphone in class and wished he could program. He said that was around 2008, making it almost twenty years earlier. The host said this explained why Masad could come up with future ideas simply by sitting down and thinking, since his mind had been working on the problem for twenty years.

Masad said he had been trying to build Replit, before it had that name, for years. The host recalled a 2024 demo in which Masad had shown him code that Masad described as object code, a reaction that Masad said he had felt as a sign that much of his decades of work was becoming obsolete. Masad said the reaction he received from engineers was controversial, and he predicted that everyone in tech would have to go through denial, depression, and eventually acceptance.

Masad said that while Claude Code partly showed the idea, he thought OpenClaw was the first time people widely experienced an LLM running a task all the way to completion, where you give it a task and it keeps going until it is done. He noted that Replit had been doing this for a year and a half by then. For him, the form factor of OpenClaw, including the ability to text it, mattered a great deal, and he said he had underestimated the power of a form factor for everyday people.

Masad said this lesson had not yet changed Replit's product. He expects it to influence future work, and he said he now believes a team needs to tinker considerably with how people experience a product, rather than assuming a clear vision will be obvious to users. He also described how he had previously avoided watching users, and how his colleague forced him to do so early on. He said it was mortifying to watch users confused by the software or hitting bugs, and that the experience helped build a company culture focused on staying close to users.

Masad said Replit's conceptual understanding of its product has always included hard, mechanical engineering, even though the product looks simple. He said the company works very hard to make the user experience simple, while the back-end engineering is very ugly. He described systems programming as part of the company's DNA.

Masad said many programmers dismiss Replit because it looks simple, but people who later join the company are often surprised by how much work sits behind the scenes. He said Replit does system administration so that users do not have to. He argued this depth helps the company win enterprise customers, who are more sensitive to reliability and security, while many competitors take reckless approaches to building insecure apps. He said the company adds guardrails that sometimes make its product slower or more expensive, but enterprises like them.

Asked whether the work done to make Replit collaborative in its first version also helped with vibe coding, Masad said yes, and described how plug-and-play the integration was. When an agent is placed into a Repl, he said, you are collaborating with the agent. In the initial version, the agent was essentially another user on the same system, and it could work on the same project as a human without special handling.

Masad said this made Replit well positioned for vibe coding because it already had the other pieces. He described how the company was organized into teams for code execution and sandboxing, deployment and cloud services, the glue between infrastructure and the IDE, and the IDE itself, and that it then created an agent team in the middle. He said infrastructure teams were now exposing tools for agents, and IDE teams were building edit tools for agents rather than interfaces for humans.

The host asked whether an IDE for agents had existed, and Masad explained that agents now generate code and can simply be given a shell. He said agents can write shell commands and do everything with basic commands like cat. He added that agents are now very good at this, but the team needed an IDE early on because the agents were bad at the task.

Masad said that editing code was once a hard problem, since agents sometimes produced mountains of text that had to be detected and looped over. Replit needed that early tooling to be among the first to offer the feature, and he said it would not have been possible to launch much earlier. He noted that the first version worked only about 70% of the time, and that customers accepted this because the experience felt magical and was an improvement over copying code back and forth from ChatGPT. Many early paying customers were people who would otherwise have hired outsourced programmers, such as a $100,000 app built by a freelance engineer.

Masad said Replit's Bounties grew quickly, with people who used Replit to write bounties already using AI to write the work. However, when he looked at the trend, he concluded that if this was where the future was heading, the bounty model would be a dead business. Replit shut it down because he expected people would eventually ask agents to do the entire task, without going through an intermediary.

Masad described enterprise use as already happening. Large companies, including banks, often pay millions for SaaS or legacy software, and they find a few vibe coders inside the company who have adopted Replit and begin replacing those systems, starting with internal work. He gave an example in which a company had scoped a project at $10 million but spent $250,000 on Replit. He said the outsourcing industry is being affected, and that Replit's biggest impact may be in companies that do not describe themselves as tech companies, such as banks and manufacturers, which he said need something like Replit or risk harming themselves.

Masad said that enterprise readiness requires giving administrators and IT departments a great deal of control over what people can do, as well as setting budgets and rules. He described this as a kind of sandboxing for humans. He framed it as the principal-agent problem, since end users are not the decision makers for the buyer, which he said is why much of the history of enterprise software has been poor.

Asked how to balance giving people what they want against the need for security, Masad said the question was tough. He disagreed with rationing tokens, which some companies handle by giving each person a monthly budget, and said companies should instead focus on return on investment. He said Replit's customers show that one person can drive a hundred times the output of the next, and that agents may make this gap larger. He said the company may build tools that help companies see where valuable work is happening.

Masad said the company is fixing one of the biggest problems inside big companies: they cannot figure out the value of work done by individual people. He said that if a product makes organizations even 10% better at that, it addresses the fundamental problem of large companies. He also said this is much bigger than vibe coding.

Masad said Replit no longer describes itself as a vibe coding company. He pointed to a blog post titled The Self-Driving Company, arguing that early adopters of technology, especially AI, are working differently, reflecting a fundamental shift in how work is done. He said Replit, as a company, is its own customer zero and serves as a lab for figuring out how people doing valuable work can be identified.

Masad said that many of the things Replit has popularized have been copied. He cited the idea of going from idea to app quickly, which he said nearly every startup now uses as its pitch. He said the competitors are copying an older version of Replit, comparing it to a movie in which they copy a frame from six months earlier. He argued Replit needs to keep moving, and that the goal is for Replit to be in every company that has competitors.

He said the company should not rely on a big strategy announced in advance. Instead, the company has a general goal and pursues it incrementally, with customers telling the story. He described internal changes in a similar way, saying that what felt discontinuous to the outside world, such as the agent, was incremental internally.

The host said Replit has been living in the future more than other companies, and Masad agreed, adding that this is more true than AI specifically. He said it is often frustrating, because things do not work yet and people change very slowly. He gave the example of cloud adoption, and asked whether people were convinced about it, which he suggested was still not the case.

Masad described an internal prediction that people will move from one agent to ten, then a hundred, then ten thousand. He said ten thousand agents cannot run on a laptop, which would eventually push work away from local machines. He noted that people still do most of their work on their desks, and suggested the scale of agents might finally convince them to move to the cloud.

Masad described a hackathon with a very large telecommunications company. Employees took a survey before and after using Replit, which asked how anxious they were about AI and how worried they were. Before using Replit for about an hour, about 80% said they were anxious and worried about their jobs. After using it, the answers flipped: about 80% were optimistic and 20% were still worried.

Masad said people in Silicon Valley constantly underestimate how far the rest of the world is from what has already been experienced there, and that it will take a long time for people to catch up. He framed this as a moral issue, saying the mission is to help these companies transform and to get people off the beach and onto the surfboard.

Masad argued that if companies are only two years behind, that is forever behind in AI terms, since vibe coding did not exist two years ago. He said most banks are using something like a 2023-era ChatGPT model, which is barely connected to any systems and so is not very useful. Silicon Valley, he said, has already moved on to agents that do things on a user's behalf, while most companies are still working out what to do with legacy software they want to rewrite.

The host noted that companies cannot simply ignore legacy systems. Masad said early on, companies asked Replit to rewrite their COBOL code, and his team refused to write the code for them, preferring to teach them how to use the tool. He said there are now AI companies that will rewrite COBOL for customers.

Asked how he stays ahead, Masad said most people treat things as static, while his default assumption is that change will happen exponentially. He said that as a child reading science fiction and about careers, he developed two things: the ability to accept change and be comfortable with it, and the expectation that computers will do more and more of the things in his skill. He said he tries to expect this and build toward it, and to skate to where the puck is going.

The host said that someone who assumes dramatic change is well suited to a world changing every six months. Masad responded that he feels more in his element now, and the host added that this is part of why Replit is doing well.