Replit CEO Amjad Masad says AI coding agents will scale from one per user to 10,000 -- and the winners are infrastructure companies -- BigGo Finance

Replit CEO Amjad Masad has a warning for anyone still processing the idea that a machine can write working code: you are not finished yet. "I think with AI, everyone will have to go through this period of denial and then depression and then eventually acceptance," he told Paul Graham on the Y Combinator Startup Podcast. Masad says he personally grieved the loss of programming as a craft around 2020 -- years before the demo that made his company, and eventually the entire tech industry, synonymous with the phrase "vibe coding."

The striking part is not that Masad saw it coming. It is why Replit, a small cloud coding platform, beat the world's best-funded AI labs to a market they were better positioned to attack. His answer is a lesson in where advantage actually lives in the AI economy -- not in the model, but in the machinery an autonomous agent needs when it is done thinking and has to actually run a program. That distinction matters for investors trying to figure out which companies capture lasting value as AI capability floods the market.

How Replit shipped agents before the model companies did

Masad traces the idea back to 2008, when he wanted to program on a Symbian phone. Replit began as a collaborative coding environment in the browser, then spent years building components that seemed mundane at the time: code execution, sandboxing, cloud deployment, databases. When large language models advanced enough to drive agents, Replit already had the pieces. "When you put an agent into a repl, it's like you're collaborating with the agent. The agent is another user," he said.

The signal came from watching users who had never programmed before. Masad invited some to the office and saw them copying code from a ChatGPT window into Replit, then copying errors back into ChatGPT, then pasting the fixes in -- a slow, manual loop. The insight was to automate the whole loop with a single goal: "the program is working." He said the model companies were surprised Replit's approach worked at all, and that in 2024 the engineering harness needed to make an agent function was roughly 10 times more complex than it is today.

Graham describes the July 2024 demo in terms that capture the moment. "Replit's like this iceberg where there's this visible bit above the water that's like look you just tell it in English what to do and it writes your code and then underneath there's all this stuff for deployment and sandboxing," he said. The demo, Graham admitted, made Masad look like a Bond villain. Masad concedes the outfit -- black shirt, not a turtleneck -- did not help.

The agent failed 30% of the time, and users did not care

The launch in September 2024 was far from polished. Graham guessed the early agent failed half the time. Masad corrected him: "It's about 30%." Users tolerated it because the alternative -- shuttling code between chat windows -- also failed, and because non-programmers suddenly succeeded most of the time. Andrej Karpathy called it a "feeling of AGI" moment.

Before the agent, Replit already had a signal most people missed. Autocomplete alone accounted for 50-60% of the code generated on the platform. The company had been an AI coding business long before it had a name for it.

What is underneath the iceberg

Masad's moat is not glamorous. Replit built its own file system from scratch so that an agent's changes are transactional and fully reversible -- in plain terms, if the agent breaks something, you can always undo it. The company employs kernel and operating-system engineers. Its cloud services let an agent provision databases and deploy real infrastructure through tool calls, not screenshots of instructions.

That depth is winning enterprise deals, Masad argues, because large buyers care about reliability and security. He says many competitors do "really reckless things" with security when agents write code, while Replit's guardrails make the product "sometimes slower, sometimes more expensive" -- a trade-off enterprises accept. The candid admission embedded in that pitch is that Replit is deliberately choosing to be slower and pricier than newer entrants, betting that trust outlasts speed.

The early technical war was uglier than the current product suggests. In 2024, foundation models had no native edit tool, so Replit built one. Models hallucinated line numbers, forcing Replit to make the edit tool "ridiculously robust." Models also got lazy, writing a function and stopping halfway with a note to finish later, producing code that would not run. Masad says both OpenAI and Anthropic had the same bug, which he attributes to chat-model training that rewarded stopping early rather than producing mountains of text. Replit had to detect the pattern and loop. Today an agent can simply be given a shell and will use basic commands to do everything, making the elaborate IDE-for-agents layer unnecessary -- a reminder that each generation of models erases someone's engineering advantage.

The business model nobody expected: scrappy insiders replacing outsourcing

Replit's first paying customers were people who would otherwise hire outsourced developers. The economics explain why. Graham said he initially thought outsourcing shops were a terrible business -- and then realized his mistake. "Arguably the whole existence of the outsourcing shops was evidence of a market there," he said. "I thought of them as just terrible -- but instead I should have thought, wow, that's a sign of demand."

Masad puts the gap at roughly $100,000 for an engineer-built app versus about $1,000 on Replit. Inside large companies, the buyers are not the IT department. They are ambitious non-engineering staff -- a sales lead building a tool for the sales team -- who then get promoted, which reinforces adoption. Masad cites customers replacing systems scoped at $10 million for about $250,000 using Replit. He acknowledges an identity question inside Replit about whether enterprise work fits the mission, and resolves it by framing those users as entrepreneurs rather than bureaucrats.

He also killed a product on the strength of the same logic. Replit Bounties, which paid teenage hackers on the platform to take jobs, grew fast -- but Masad shut it down. "That's the thing about the AI business today is that you have to just accept that everything that you're doing is like getting commoditized constantly and you have to be doing the next thing," he said. His forecast is blunt: human-intermediated outsourcing marketplaces are a dead business because people will ultimately ask agents for the entire thing directly.