Amjad Masad and Paul Graham on Replit and the Self-Driving Company

Replit CEO Amjad Masad tells Paul Graham how Replit built coding agents before vibe coding had a name, and why he now sells banks a self-driving company. The conversation, from Y Combinator’s How to Build the Future series, covers the 2024 launch of Replit Agent, the infrastructure that made it possible, why enterprises are replacing $10 million software projects with $250,000 of Replit, and how Masad trains himself to predict where technology is going.

TLDW

Replit got to vibe coding first because it watched non-programmers paste code back and forth between ChatGPT and Replit, then automated the whole loop: write, run, read the error, fix, repeat until the program works, and provision the database and hosting along the way. Years of systems work on collaborative coding meant the agent could simply join a project as another user. Today the growth is in large, non-technical companies, where ambitious employees use Replit to replace expensive software, and Masad’s bigger pitch is the “self-driving company,” with Replit as the way AI flows into a firm. Graham’s closing diagnosis is that Masad wins because he assumes exponential change by default, while most people treat the world as static.

Thoughts

Token rationing is the wrong control, and Masad’s objection to it may be the most commercially important thing he says. Companies are handing every employee a flat budget, say $100 of tokens a month, while his own customer data shows one person can drive 100 times the value of the next, a spread he says agents have widened from 10x to 100x or even 1,000x. Graham pushes the point further than Masad does: if Replit can show a company which projects its tokens are actually paying for, it is attacking the oldest problem big firms have, which is that they cannot see the value of work done by individual people. Make them even 10% better at that and you have fixed something far larger than app building. It also explains Masad’s tactic of giving IT departments the budgets and controls they ask for now, as the way in, and inventing the better measure later.

The outsourcing argument is the clearest account in the conversation of who actually pays for AI coding. Replit’s first big-spending customers were not developers. They were people who would otherwise have hired an outsourced programmer, and the choice was roughly $100,000 for a contractor against $1,000 for Replit, a two-order-of-magnitude gap Masad says still holds. Graham admits he used to see outsourcing shops as sellers of terrible software and should have read them as proof of demand. The same logic killed Replit Bounties, the marketplace where teenage hackers built software for clients with Replit’s early AI: once the client could ask the agent directly, the middle layer was a dead business. The enterprise version is the claim that a bank scoped a project at $10 million and finished it for $250,000. That figure comes from Masad, not an audit, but the direction matters for anyone selling a service an agent can now do.

Replit’s head start looks less like foresight and more like having built the right plumbing for a different reason. Years of work on collaborative coding meant an agent could join a project as just another user, and the teams building sandboxes, databases, deployment and the IDE for people simply turned to building tools for the agent. Masad is candid that this edge erodes. The 2024 harness was about ten times more complex than today’s, because weak models needed a custom edit tool, hallucinated line numbers and got lazy halfway through a function, and every gain in model capability makes the agent layer easier to copy. What does not get easier to copy is the part below the waterline: a transactional, fully reversible file system, kernel engineers, and security guardrails that make the product slower and pricier but are exactly what enterprises buy.

The most hopeful data point is also the strangest. In a hackathon at a large telecom company, about 80% of staff said they were anxious about AI before an hour on Replit, and about 80% said they were optimistic after it. One hour is not a study, and a vendor’s own hackathon is not a neutral setting, but it fits the larger claim that most of the world has not yet felt what Silicon Valley has. Masad says most banks, outside engineering, run something like 2023 ChatGPT, an LLM barely connected to anything. Graham sharpens it: two years behind is forever behind when vibe coding did not exist two years ago. His advice for children works for companies too. When a tidal wave approaches, the safest place is on a surfboard, not on the beach holding a palm tree.

The closing answer on how Masad “lives in the future” is less mystical than it sounds. He makes predictions, revises them, and anchors them to Ray Kurzweil’s chart of world compute on a log scale, a straight line he says AI did not bend. The compute brought on AI, not the other way round, and neural networks that failed in the 1980s worked once they had billions of times the resources. That is how he looked at GPT-2 in 2019 and pictured chat, agents and personal assistants. Survivorship bias applies, since we hear from the founder whose bets paid off. But Graham’s summary is the habit worth copying: most people implicitly treat the world as static, and Masad’s default is that it will change exponentially. That is a cheap assumption to adopt and an expensive one to ignore.

Key Takeaways

Before Replit Agent, Replit users were already generating something like 50 to 60% of their code with AI autocomplete, so full vibe coding felt incremental inside the company.
The idea came from watching non-programmers in the office copy code from ChatGPT into Replit, copy the error back, and paste the fix in again. Replit automated that loop.
Replit showed two things the model labs had not: an LLM can run in a loop until the program works, and AI can provision real infrastructure such as databases and hosting through tool calls. Masad says the model companies reached out to ask how.
Graham saw the agent in July 2024 and it launched in September 2024. His reaction was dismay: Masad told him not to look at the code because it was “object code.”
Replit’s 2015 to 2016 pitch deck predicted training a model on millions of users’ code to generate programs, before the Transformer paper.
Masad credits OpenClaw, more than Claude Code, with convincing the wider world at the end of 2025 that agents can run tasks to completion, and says he underestimated the power of a form factor you can simply text.
The agent was first built as “just another user” on Replit’s collaborative system, and the infrastructure, cloud services and IDE teams turned to building tools for it.
Early agents worked maybe 70% of the time. Users tolerated it because the alternative was paying an outsourced developer about $100,000 instead of about $1,000 on Replit.
In enterprises, a few internal vibe coders replace costly SaaS and legacy systems, starting with internal tools, and often get promoted for it.
Masad opposes flat token rationing. He argues companies should measure ROI, since value per person varies by 100x or more with agents.
The “self-driving company” is Replit’s new framing: Replit is customer zero, and the platform is pitched as the way AI flows into firms that cannot afford to sit still.
Replit’s internal prediction is that people go from one agent to 10, 100 and eventually 10,000, which cannot run on a laptop and will push work into the cloud.
Masad’s forecasting method: make predictions, revise them, and trust the long exponential trend in compute rather than the instinct that things will stay the same.

Chapters

0:47 Vibe Coding Before It Had a Name

Graham opens by saying Replit basically invented vibe coding in 2024, which already feels like a technological age ago. Masad says it was incremental for Replit, following the YC habit of watching what users do. With autocomplete, users were already generating 50 to 60% of their code, so the leap to telling the machine what to do in English was smaller than it looked from outside.

2:43 Watching Users Copy-Paste From ChatGPT

Masad invited users into the office and watched people who had never coded shuttle code and errors between ChatGPT and Replit. Automating that meant understanding errors and looping until the program works, rather than taking one turn the way chat tools did. Replit also showed that AI can provision infrastructure, because it had years of database and hosting systems to expose as tools. Andrej Karpathy called the result a “feel the AGI” moment, and Graham admits it freaked him out.

5:27 The Origins of Replit and Predicting the Future

Graham recalls his own principle from first backing Replit: make something 10% easier and you get twice as many users, though he “didn’t mean make it this easy.” Masad describes a 2015 to 2016 deck predicting machine-learned code generation and Replit’s early, much-mocked mobile app. He traces the idea back to 2008, wishing he could program on his Symbian phone in class. Graham’s point is that “just sitting down and thinking” works only after 20 years of obsessing over one problem.

10:27 Grieving the Craft and Obsessing Over Users

Masad says everyone goes through denial, depression and acceptance with AI, and that he grieved the craft of programming around 2020. He thinks the wider world was convinced at the end of 2025, mostly by OpenClaw and partly by Claude Code. He also admits he hated watching users struggle with the software, and credits co-founder Haya Odeh with forcing a culture of constant user contact and giving Replit its distinctive look.

15:13 Why Enterprises Pick Replit, and the Agent as Another User

Masad describes Replit as an iceberg: a simple English prompt on top, and sandboxing, deployment, kernel work and a transactional, reversible file system underneath. Guardrails make the product slower and costlier, which enterprises like and some reckless competitors skip. Because Replit was built for collaboration, the first agent was simply another user, and every team reoriented to building tools for it, including a custom edit tool for models that hallucinated line numbers or left “finish later” comments in the code.

20:17 It Barely Worked, and Users Loved It

At launch the agent failed maybe 30% of the time, but for someone who could not code at all, working 70% of the time was magic. The first heavy spenders were people who would otherwise pay outsourcing shops, a gap of about $100,000 against $1,000. Graham concedes the outsourcing industry was a demand signal he dismissed. Masad shut down Replit Bounties because agents would eventually remove the human middle layer altogether.

23:13 $10 Million Projects Now for $250,000

Inside banks, manufacturers and other large companies, a few internal vibe coders are replacing expensive SaaS and old systems, starting with internal tools. Masad tells his team these ambitious employees are entrepreneurs, so enterprise work fits the mission. Graham notes they are the people companies will trust when leadership decides it must “use AI.” Being enterprise-ready is a schlep: admin controls, budgets and rules for IT.

25:45 Sandboxing for Humans and the Case Against Token Rationing

Graham calls enterprise controls “sandboxing for humans.” Masad describes the principal-agent problem in enterprise software, where users are not the buyers, and argues against flat token budgets because value per person varies by 100x or more. Graham says helping companies see who is doing valuable work would fix the fundamental problem with big companies. Masad’s plan is to give customers what they ask for first, then invent something better.

28:14 The Self-Driving Company

Masad says Replit no longer describes itself as vibe coding. Its “self-driving company” idea treats Replit as customer zero for a new way firms are organized, with Replit as the platform for that shift. Graham says the best startups reinvent themselves like new companies while already big, and Masad argues AI’s commoditization makes that mandatory. Competitors copying “idea to app fast” are, in Graham’s phrase, copying a frame from a movie six months old.

33:28 Steve Jobs, 10,000 Agents and Killing the Sacred Cow

Steps that look sudden to outsiders are incremental to insiders, as with Steve Jobs wanting a tablet long before the technology allowed it. Replit predicts people will go from one agent to 10,000, which forces work off laptops and into the cloud. Masad says he can throw away what he built when computers get good at it, and points people to Graham’s essay on keeping your identity small. A telecom hackathon flipped staff from 80% anxious to 80% optimistic about AI in one hour.

38:25 Two Years Behind, and the Compute That Brought On AI

Most banks run something like 2023 ChatGPT outside engineering, while legacy COBOL is finally being rewritten. Asked how he lives in the future, Masad cites Ray Kurzweil’s straight line of world compute and a line he saw on X: the compute brought on AI. Graham concludes that Masad’s default assumption is exponential change, which makes him perfectly suited to a world that shifts every six months.

Notable Quotes

“I think with AI, everyone will have to go through this period of denial and then depression and then eventually acceptance.”
Amjad Masad, on how engineers come to terms with AI writing code

“That’s true for anything consumers use. If you make it 10% easier, you don’t just get 10% more people, you get twice as many.”
Paul Graham, on the principle behind his early bet on Replit

“Arguably the whole existence of the outsourcing shops was evidence of a market there.”
Paul Graham, on who Replit’s first paying customers really were

“This is why big companies are the way they are. They can’t figure out the value of stuff done by individual people.”
Paul Graham, on why measuring token ROI could matter more than vibe coding

“If you have this tidal wave approaching the beach, the safest place to be is surfing on it. Not standing on the beach holding on to a palm tree.”
Paul Graham, on what to teach children about AI

“If they’re just two years behind, that’s forever behind, right? Cuz two years ago, vibe coding didn’t exist.”
Paul Graham, on how far most companies lag Silicon Valley

“It’s not that AI brought on the compute. The compute brought on AI.”
Amjad Masad, quoting a tweet that sums up his forecasting method

Watch the full conversation between Paul Graham and Amjad Masad here.