Amjad Masad and Paul Graham on How To Build The Future: Key Points

A summary of Paul Graham's interview with Replit CEO Amjad Masad: how Replit Agent came about, enterprise use, the self-driving company and living in the future.

▶ Watch the full interview on YouTubeY Combinator · 7 November 2026

Paul Graham interviewed Amjad Masad, the co-founder and CEO of Replit, for How To Build The Future, a series on Y Combinator's YouTube channel. The episode was published on 7 November 2026 and runs about 44 minutes^[1]. It had more than 82,000 views by 12 November 2026. Graham backed Replit early, and much of the talk is about how Replit came to build a coding agent before most companies did, what large companies now use it for, and why Masad thinks he can predict where technology is going. This page summarises what was said, grouped by topic, with timestamps.

Topics and timestamps

The topic names are taken from the chapters in the episode description, some of them shortened^[1].

How Replit Agent started

Graham opened by saying Replit basically invented vibe coding in 2024. Masad said that for Replit it was incremental. Before the agent, he said, Replit's AI autocomplete was already writing perhaps 50 to 60% of users' code.

The idea for the agent came from watching users. Masad said he invited some users in and watched people who had never coded open ChatGPT in one window and Replit in another. They copied code into Replit, copied the error back into ChatGPT, and pasted in the fix. Replit set out to automate the whole process. Graham pointed out that this meant the system had to understand the errors, not only pass them back and forth. Masad added that it had to keep looping until the program worked, where ChatGPT and most AI tools of the time took one turn.

Masad said Replit showed two things. The first was that a language model can run to completion. He said the model companies contacted Replit because they were surprised the agent worked, and that the harness around the model in 2024 was 10 times more complex than it is today, because the models were less capable. The second was that AI can provision infrastructure, such as spinning up a database. Replit could do this because it had spent years building its own infrastructure.

Graham saw the agent in July 2024, and Masad said it launched in September 2024. He said Andrej Karpathy called it a "feel the AGI" moment. Graham said it "freaked me out".

The origins of Replit, and predicting the future

Masad said Graham gave him the principle behind Replit when Graham first backed the company: there is a super-linear relationship between how easy programming is and how many people program. Graham agreed. He said that if you make something 10% easier, you get twice as many people, not 10% more. Graham joked that he did not mean for it to become this easy.

Masad said Replit's original document, from around 2015 or 2016, described training a machine learning model on the code written by millions of people on the platform to predict code. Graham noted this was before the Transformer paper. Masad said Replit was also the first programming platform with a mobile app, and that people laughed at the idea of programming on a phone. He first wished he could program on his phone in 2008, in class, with a Symbian smartphone.

Masad said people underestimate their ability to predict where the future is headed and build towards it, and that few people sit down and think. Graham's answer was that this works for Masad because he had been thinking about the same problem for about 20 years.

The Bond villain demo and grieving the craft

Graham described the demo in the summer of 2024 as being shown something frightening. He remembered Masad telling him not to look at the code, because it was object code, and his reaction that the code he had spent decades writing was now object code. Masad recalled that Graham called him a Bond villain. Graham said his own style of programming is the opposite: he will spend three years writing a thousand lines of code and be glad to cut it to 980.

Masad said everyone will go through denial, depression and then acceptance with AI. He went through it around 2020, so he had already grieved the loss of the craft he loved. In the conversation, the end of 2025 was named as the point when most people were convinced that AI would write all the code. For the wider world, he credited OpenClaw more than Claude Code, because it was the first time many people saw an AI run a task to completion, in a form factor they could text. He said he had underestimated how much the form factor matters for everyday people.

Watching users, and Replit's character

Masad said he did not like watching users, because it was painful to see them confused or hitting bugs, and that his co-founder Haya pushed him to do it early on^[2]. He credited her with making talking to users part of the company culture. Graham praised Replit's early graphic design, and Masad said it helped with the first audience, many of whom were teenage hackers.

Masad said Replit works hard to keep the interface simple while a lot of difficult engineering happens behind it, so some programmers dismiss it as simple. Graham compared Replit to an iceberg: the visible part is telling it in English what to do, and underneath is deployment and sandboxing. Masad said the company employs kernel and operating system engineers and built its own file system from scratch to be transactional and fully reversible.

The agent was just another user

Masad said the work Replit had done for collaborative coding carried over to the agent almost directly. In the first version, the agent was just another user on the same system, so a person and an agent could work on the same project. The company's teams also fitted: infrastructure, cloud services and deployment, the layer between them and the editor, and the editor itself. Replit created an agent team and put it in the middle, and the other teams began building tools for the agent.

In 2024, Masad said, even editing code was hard. Models had no native edit tool, they hallucinated line numbers, and they got lazy and left notes to finish a function later. He said OpenAI's and Anthropic's models both had that last problem, and he thinks it came from training chat models to stop early. Replit had to detect it and loop. Today, he said, you can give an agent a shell and it can do everything with basic shell commands.

It barely worked, and users loved it

Masad said the agent launched in beta and failed about 30% of the time. Users put up with it because it was a magical experience. Graham added that for someone who cannot program at all, software that works 70% of the time is a big improvement.

The first customers who paid Replit a lot of money, Masad said, were people who would otherwise hire outsourced programmers. Instead of paying about $100,000 for an app, they paid Replit about $1,000, a difference of two orders of magnitude that he said still holds. Graham said he had seen outsourcing shops as terrible, but should have seen them as a sign of demand. Masad said Replit had earlier run Replit Bounties, where teenage hackers on the platform built software for others with help from Replit's early AI. It grew fast, but Replit shut it down because people would eventually ask agents for the whole thing.

Enterprise customers and $250,000 projects

Masad said Replit now goes into large multinational companies and banks that pay millions for SaaS or old software. They find a few people inside who have adopted Replit and are replacing systems, starting with internal tools. Some told Replit that a project they had scoped at $10 million cost them $250,000 on Replit. He said the outsourcing industry is "sort of dead at this point". Tech companies are also customers, but he thinks Replit has the biggest impact on companies that do not see themselves as tech companies.

He said enterprise customers care more about reliability and security than consumers do. Replit adds guardrails, which sometimes make it slower or more expensive, and he said many of Replit's competitors do reckless things and build insecure apps. Masad said a lot of the work of being enterprise-ready is giving IT and administrators controls and budgets. Graham called this sandboxing for humans. Masad said Replit's growth in enterprises caused an identity question inside the company. His answer to his team was that ambitious people building things inside companies are core to Replit's mission, and that people who bring Replit into their organisation often get promoted.

Masad said he disagrees with companies rationing tokens, for example giving every employee $100 of tokens a month. Graham said companies should care about return on investment, and Masad agreed. He said Replit sees a power law inside companies, where one person can produce 100 times what the next one does. In the short term Replit gives companies the controls they ask for, and then it wants to build something better. Graham said a tool that showed which projects are most valuable would fix the biggest problem big companies have.

The self-driving company

Masad said Replit no longer describes itself as vibe coding. It wrote a blog post called The Self-Driving Company^[3], about a shift in how early adopters of AI work and how firms will be organised. Replit is customer zero and uses itself as the lab. He said Replit sees itself as "the platform for the transformation for the self-driving companies". He had always known that making software creation more accessible would have a big impact, but had not appreciated that it would change how people work. Graham summed up the idea as Replit being the way AI flows into companies.

Masad said AI is a massive commoditisation force, so reinventing the company is now close to mandatory. Everything Replit popularised has been copied, he said, such as going from idea to app quickly. Graham said copycats are copying a frame of Replit from six months ago. Masad said founders should not announce a big new strategy but should let customers tell the story, and Graham described this as having a general goal and getting there in increments.

On invention, Masad said changes that look sudden from outside are often incremental inside a company, and gave the example of Steve Jobs, who tried to make an iPad in the 1980s.

10,000 agents and the rest of the world

Masad said Replit's internal prediction is that people go from one agent to 10, to 100, and eventually to 10,000 agents. Ten thousand agents cannot run on a MacBook, and Masad expects work to move away from local machines, although most people still work on their desktops. Graham said this may finally convince people that this kind of work has to live in the cloud.

Masad said he can let go of skills once computers get good at them, and has started over many times in his career. He pointed people to Graham's essay Keep Your Identity Small. When parents ask him what to teach children, he says flexibility, adaptability and lifelong learning. Graham's own answer is that when a tidal wave is coming, the safest place is surfing on it, not holding onto a palm tree on the beach.

Masad described a hackathon with a very large telecommunications company. Before an hour of using Replit, he said, about 80% of participants were anxious about AI and their jobs. Afterwards about 80% were optimistic. He said Silicon Valley underestimates how far behind the rest of the world is: most banks, outside engineering, use something like the ChatGPT of 2023, an LLM barely connected to anything. Graham said two years behind is a very long time when vibe coding did not exist two years ago. Masad said legacy COBOL code is starting to be rewritten. When companies once asked Replit to rewrite their COBOL, he told them Replit would teach them to use the product instead, and now other AI companies offer that service.

How Masad tries to live in the future

Asked how he lives in the future, Masad said he actively makes predictions, revises them and checks them, and thinks it is a skill you can get good at. Early on he read Ray Kurzweil, whose log-linear chart of the world's compute is a straight line that, Masad said, AI has not bent. He quoted a tweet saying "It's not that AI brought on the compute. The compute brought on AI." Graham added that neural networks were thought not to work in the 1980s, and that they started to work when given billions of times more resources.

Masad said that when he saw GPT-2 in 2019 he imagined chat, agents and personal assistants. He said most people have an instinct to list all the reasons something will not work. Graham said the difference is that Masad's default assumption is exponential change, and that this makes him suited to a world where everything changes every six months.

Sources

1. How To Build The Future: Paul Graham & Amjad Masad — Y Combinator, YouTube (7 November 2026; views checked on 12 November 2026)
2. Replit: About Us, which lists Haya Odeh as co-founder
3. Replit blog: The Self-Driving Company