How Replit Invented Vibe Coding: Amjad Masad's Origin Story
Replit CEO Amjad Masad told Paul Graham how watching users paste ChatGPT code into Replit led to the AI agent that sparked vibe coding.
What is vibe coding and where did the term come from?
Vibe coding means describing what you want in plain English and letting an AI agent write, run, and fix the code until it works, without you reading or writing the code yourself. The practice traces back to Replit, whose CEO Amjad Masad described its origin in a conversation with Paul Graham. According to Masad, the breakthrough wasn’t a sudden idea. It came from watching Replit users manually shuttle code between ChatGPT and the Replit editor, then building a product that automated that exact workflow into an AI agent that could run a coding task to completion.
TL;DR
- Replit’s agent grew out of a manual workflow Masad observed firsthand: users copying code from ChatGPT into Replit, pasting back errors, and repeating the cycle by hand.
- Before the agent existed, Replit’s autocomplete tools were already generating roughly 50 to 60 percent of code for some users, which gave the team confidence that automating the rest was possible.
- The hard engineering problem wasn’t generating code, it was looping the model until the program actually worked, meaning the system had to read errors, retry, and only stop when the task was genuinely complete.
- Replit demonstrated two things that were new at the time: running an LLM to completion on a multi-step task, and letting AI provision real infrastructure like databases, not just write snippets.
- The public demo in July 2024 and launch in September 2024 reportedly prompted outreach from AI model providers who were surprised their own models could sustain that kind of looped, autonomous workflow.
- Masad says the engineering harness required to make 2024-era models act as agents was far more complex than what’s needed today, because as underlying models got more capable, less custom scaffolding was needed around them.
- Masad frames the broader shift as emotional as much as technical: he says programmers, including Graham, go through denial, grief, and eventual acceptance as AI takes over the actual writing of code.
Nothing gets lost.
How did Replit actually discover the vibe coding workflow?
Masad describes it as incremental rather than a flash of insight. Replit had been building autocomplete and AI assistance features for a while, and by his account these tools were already writing somewhere around half to two-thirds of the code for users before any “vibe coding” product existed. That gave the team a baseline: if AI could already produce most of a codebase with a human stitching pieces together, the next question was whether a system could handle the stitching too.
The clearer signal came from direct user observation. Masad says he invited users to Replit’s office and watched how people with little or no coding background were already building full projects using AI, just not inside any integrated tool. Their process was manual: open ChatGPT in one window, open Replit in another, write a prompt, copy the generated code into Replit, run it, copy any error message back into ChatGPT, get a fix, and paste it in again. Masad called it a loop that users were running by hand, over and over.
That observation reframed the problem. It wasn’t enough to connect Replit to a model through an API. The system needed to automate the entire loop: generate code, run it, interpret the resulting error, decide on a fix, and repeat, with the loop only ending when the program worked rather than after a fixed number of exchanges.
What made building an AI agent that runs to completion so hard?
At the time Replit built this, large language models typically handled only a couple of tool calls before stopping or losing the thread. Most AI products, including early chat interfaces, worked in single turns: you asked, it answered, and any follow-up was a new, separate exchange.
Replit needed something different. The system had to keep looping indefinitely, feeding errors back into the model, until the terminal condition was “the program works,” not “the model responded once.” Understanding and acting on error output was, in Masad’s telling, the genuinely difficult part, more so than generating the original code.
Replit also pushed the agent beyond writing application code. The system could call tools to provision real infrastructure, spinning up a database or a caching layer as part of completing a task. Masad attributes this to years spent building Replit’s own underlying systems software, which meant the company had actual infrastructure to hand to an agent rather than code generation sitting on top of someone else’s platform.
Masad notes that the engineering harness needed to make 2024 era models behave this way was far more elaborate than what’s required today. As base models have grown more capable on their own, the amount of custom scaffolding needed to turn them into reliable agents has gone down, which he acknowledges also lowers the barrier for competitors to build similar systems.
How did the industry react when Replit showed this off?
Nothing gets lost.
Replit demoed the capability in July 2024 and launched it in September 2024. Masad says the reaction from some engineers, including AI researcher Andrej Karpathy, treated it as a notable moment, with Karpathy reportedly describing it in terms comparable to a glimpse of more general AI capability, since it showed language models completing open-ended, multi-step software tasks rather than answering single prompts.
Masad also says model providers themselves reached out, surprised that their models could sustain the kind of looping, self-correcting behavior Replit had built around them. That reaction is notable because it suggests the capability wasn’t something the model builders had specifically designed for or anticipated; it emerged from how Replit structured the surrounding agent loop.
Graham’s own reaction, by his account in the conversation, was closer to unease. He describes watching the demo and being told not to look at the underlying code, calling it “object code,” a framing that stuck with him because it implied the code itself, the thing he’d spent decades caring about, was no longer the point.
Is vibe coding just a product feature, or a bigger shift?
Masad resists describing Replit purely as “the vibe coding company,” framing the broader goal instead as building a platform for AI driven software creation generally, extending toward what he calls self-driving companies. But he’s direct that the vibe coding workflow came from a specific technical insight: that an AI agent could be looped to completion and trusted to provision real infrastructure, not just emit text.
He also connects this to a long standing thesis of Graham’s, voiced in the conversation: that making programming easier doesn’t produce a proportional increase in the number of people who program, it produces a much larger one. Masad says that’s part of why Replit pushed in this direction rather than treating AI assistance as a minor convenience feature.
He’s candid that the emotional reception has been mixed, including his own reaction to watching Replit’s agent in action the first time, which he describes as feeling genuinely strange given his own history as a hands-on programmer. He frames this as a pattern he expects most programmers to go through: resistance, then something like grief over craft, then eventual acceptance.
Frequently Asked Questions
Who invented vibe coding?
Based on this account, Replit and CEO Amjad Masad are credited with building the first AI coding agent that could run a software task to completion, autonomously looping through errors and fixes rather than requiring a human to manually relay information between a chat model and a code editor.
When did Replit launch its vibe coding agent?
Masad says Replit demoed the capability in July 2024 and launched it publicly in September 2024.
What’s the difference between autocomplete and vibe coding?
Autocomplete suggests code as a human writes it, but the human still assembles, runs, and debugs the project. Vibe coding, as described here, hands the entire loop to an agent: writing code, running it, reading errors, and retrying until the task is actually finished.
Why was looping to completion technically difficult in 2024?
Models at the time generally handled only a few tool calls reliably and most AI products worked in single turns. Building a system that kept looping based on real program errors, stopping only once the software worked, required a large amount of custom engineering around the model itself.
Did AI model companies know their models could do this?
According to Masad, model providers reached out to Replit after the launch because they were surprised their own models could sustain the kind of extended, self-correcting agent loop Replit had built.
