Replit was building toward the AI coding era years before it had a name. In 2024, the company shipped an agent that could take a request in plain English, work through errors, provision its own infrastructure, and keep going until the software actually worked. The model companies even reached... out to ask how Replit had pulled it off. Co-founder and CEO Amjad Masad had been thinking about making programming radically more accessible since 2008, and Replit’s original pitch deck predicted AI-generated code before the Transformer paper was published. Today, that same technology is moving into banks, manufacturers, and other large companies, where Amjad says teams are replacing software projects once scoped at $10 million with $250,000 of Replit. On How to Build the Future, Paul Graham sits down with Amjad Masad to talk about why Replit ⠕ was so well positioned when coding agents arrived, what comes after vibe coding, and his bigger vision for the “self-driving company,” where AI agents increasingly build and operate the software a company needs. 00:00 - "You guys basically invented vibe coding" 02:43 - Watching users copy-paste from ChatGPT 05:27 - The origins of Replit 08:00 - People underestimate their ability to predict the future 10:27 - Grieving the craft 12:52 - A culture of obsessing over users 15:13 - Why enterprises pick Replit 17:28 - The agent was just another user 20:17 - It barely worked, and users loved it 23:13 - $10 million projects now for $250,000 25:45 - Sandboxing for humans 28:14 - "The self-driving company" 30:34 - The way AI flows into companies 33:28 - What Steve Jobs knew before the iPad 35:59 - Killing the sacred cow 38:25 - The rest of the world is two years behind 41:15 - The compute brought on AI 43:46 - The person who assumes dramatic changeshow more

I'm sure @paulg never used this. I was early on Replit. It was complete trash. An IDE harness connected to LLMs that didnt work, looped endlessly, and gummed up unit and regression tests into a mashup of worthless ascii. It was so good, they open sourced it.

I am happy to be learning from the people who invented "Vibe-Coding" @replit #georgeohan

@amasad He'll yea!

an agent that keeps going until the software actually works sounded wild in 2024. now people get mad when it stops

"Work through errors" is the exact bottleneck right now. Is the agent actually understanding the logic, or just brute-forcing stack traces until the linter stops yelling? The difference determines whether the software is actually maintainable.

for a self-driving company the software part sounds believable. i'd want to know who decides which customer to say no to, because an agent will happily build whatever the loudest customer asks for

The infrastructure piece deserves some love ❤️ Giving the agent somewhere to run its code turns errors into feedback it can act on.

the hard part of that sentence is "until the software actually worked". deciding that is still mostly a human, and every agent product is a bet on moving it.

最想听企业项目那一段：原来要花大预算的软件，哪些工作真的被省掉了。

The big structural advantage was already having sandboxed runtimes. An agent needs a tight feedback loop with real stdout and stderr to self-correct, which is hard to pull off locally.

Shipping an agent that provisions its own infra back in 2024 was early. Being early and still standing is the rare part!

Replit Agent强在环境部署一条龙，小白也能跑通，坑是遇到报错会自己绕圈修，额度烧得飞快，还顺手把能用的代码也改坏。

banks move slow but that agent angle is real tbh

Model companies treat coding like text generation, but Replit treated it as an interactive environment from day one.

@amasad @grok is Replit open source or patented?

seen this pattern with founders i work with too the ones who were "early" to AI coding weren't predicting the future they were just building what felt obvious to them, years before it was obvious to everyone else

the skipped part: when the agent provisions its own infra, you inherit its cloud bill and its mistakes. debugging someone else's terraform at 2am is still on you.

watching users copy code between tools seems like a pretty clear sign of what to build next

When the Replit agent keeps going through errors, what signals make it stop and hand control back to the user?

The cost drop is the real story. When work that once needed ten million can be attempted for a fraction of that, small teams can try ideas only large companies could afford. The new bottleneck is judgment: what is worth building, and who owns it when it breaks.

Keep going until the software actually worked. Mine stops when the context window fills

Same pattern on jobsites. An agent that retries is useful only if someone still owns the plan. Who checks the plan before the next retry?

the self-driving company is believable for the software. the harder part is the humans still in the loop: when agents produce half the work, someone has to know by evening what actually got done versus what only looks done