Context
Amjad Masad is the co-founder and CEO of Replit, a browser-based platform where people write, run and deploy software. In this episode of How to Build the Future, YC co-founder Paul Graham talks with him about how Replit shipped one of the first coding agents in 2024, why the company was so well placed to do it, and where it is heading now. The conversation matters to PMs because it is a detailed case of finding a product in observed user behavior, building a moat in unglamorous infrastructure, and selling AI tools into large companies where the buyer and the user are different people. It also covers a habit Masad credits for most of Replit's timing: making explicit predictions about the future and revising them.
The Big Idea
The companies that look like they leapt ahead are usually just building, step by step, toward a future they predicted years earlier, while watching closely what users already do.
Replit's coding agent looked like a sudden breakthrough to the outside world, but inside the company it was an incremental step from autocomplete and from watching users copy code out of ChatGPT. Masad argues that because AI commoditizes everything quickly, constantly moving to the next increment is no longer optional.
Key Insights
1. The agent came from watching users
- Before Replit shipped its agent, its autocomplete was already writing roughly 50 to 60% of users' code.
- Masad invited some users to the office and watched them work.
- People who had never coded would paste code from ChatGPT into Replit, copy the resulting error back into ChatGPT, then paste the fix back in.
- Masad saw that Replit could automate that whole loop instead of leaving users to do it by hand.
- He frames this as the YC habit of looking at what users are doing and letting that tell you what to build.
2. Loop until the program works
Automating the copy-paste loop was not just a matter of wiring up an API. The agent had to understand errors and keep looping, with the stopping condition being that the program actually works, rather than answering once like ChatGPT and most AI tools did at the time. Masad says AI model companies reached out, surprised that Replit's agent worked, and asked how Replit had made it work.
3. Your harness gets commoditized
- Masad says the agent harness Replit built in 2024 was about 10 times more complex than the one it needs today, because the models were weaker then.
- As models improve, the engineering needed to turn them into agents shrinks.
- Graham asked whether that makes it easier for competitors to catch up, and Masad said yes.
- His response is to accept that everything an AI company builds is being commoditized constantly, and to always be doing the next thing.
4. Infrastructure was the hidden advantage
- Replit's agent could provision real infrastructure, such as spinning up a database or Redis and hosting the app, because Replit had spent years building that systems software.
- Masad describes Replit as an iceberg: a simple English prompt on top, and deployment, sandboxing, kernel and operating-system engineers, and a custom transactional, fully reversible file system underneath.
- Work done to make coding collaborative turned out to be "plug and play" for agents, since the first version of the agent was simply another user on the same system.
- Even Replit's team structure fit: a new agent team sat in the middle, and the infrastructure and IDE teams started building tools for the agent instead of only for humans.
5. Magic beats reliability for non-coders
- At the 2024 beta launch, Masad estimates the agent failed about 30% of the time.
- Users tolerated this because, as Graham put it, if they could not program at all, working 70% of the time was a huge improvement.
- The first big paying customers were people who would otherwise hire outsourced programmers.
- Masad puts the gap at about $100,000 for an outsourced app versus about $1,000 on Replit, a two-order-of-magnitude difference he says still holds.
- Graham's lesson: he had dismissed outsourcing shops as terrible, but their existence was evidence of real demand.
6. Enterprise buyers are not users
Masad describes what he calls the principal actor problem: in enterprise software, the end users are not the people who decide what to buy. He thinks this is why most enterprise software has historically been bad. Users now have more say, so good products are winning in large companies, but IT and security teams can still force a product to become worse for users. Replit adds guardrails that sometimes make the product slower or more expensive, which enterprises like and consumers rarely care about.
7. Value follows a power law
- Masad disagrees with rationing tokens, for example giving every employee $100 of AI tokens a month.
- He says companies should care about ROI instead.
- Replit sees a power law of ROI inside customer companies, where one person can drive 100X value, up from roughly 10X before agents.
- Graham argues that helping a big company see which individuals are creating value would fix the fundamental problem with big companies, since they cannot measure individual contributions.
8. Most companies are years behind
- At a hackathon with a very large telecom company, a survey before using Replit found about 80% of employees were anxious about AI.
- After about one hour of using Replit, it flipped to about 80% optimistic.
- Masad says most banks outside engineering are using an AI setup at roughly the level of ChatGPT in 2023, barely connected to anything.
- His point: people in Silicon Valley underestimate how far behind everyone else is, and two years behind is "forever behind" at today's pace.
Mental Models & Frameworks
Ease of use is super-linear
Graham's principle, which he gave Masad when he first invested, is that the relationship between how easy programming is and how many programmers you get is super-linear. Graham says it applies to anything consumers use: make it 10% easier and you do not get 10% more users, you get twice as many. Masad concluded that for Replit to truly succeed it had to go "the whole way" to plain-English programming. For PMs, it is a reason to treat removing friction as a growth lever, not polish.
The iceberg product
A simple surface sits on top of deep, hard engineering. Masad notes the paradox: Replit works hard to make the user experience simple, so many programmers dismiss it as simple, while new hires are surprised by how much happens behind the scenes. The hidden depth is what makes enterprise sales possible, since enterprises care about reliability and security.
Insiders see increments, outsiders see steps
- What looks like a step change to the world is often smooth from the inside.
- Replit's agent felt discontinuous to users but was incremental internally.
- Masad compares it to Apple: Steve Jobs tried to make an iPad in the 1980s, but the technology was not ready, and the iPhone later came out of tablet work he had been thinking about for years.
- Graham adds that when a new capability appears, a company like this recognizes it as something it has been thinking about for 10 years.
Treat the world as exponential
- Masad credits reading Ray Kurzweil, whose argument is that a log-linear chart of the world's compute is a straight line.
- Masad says AI did not bend that line: it is still straight.
- The two note that neural networks were thought not to work in the 1980s, but with about 10 billion times more resources they started to work.
- Graham's summary: most people implicitly treat things as static, while Masad's default assumption is that things change exponentially.
- Masad says that seeing GPT-2 in 2019 let him imagine chat assistants, agents and personal assistants, because he extrapolated along that curve.
Predict, then revise
Masad says living in the future is partly a skill you can get good at. He actively makes predictions about where things are going, then revisits and revises them. He adds that you must go against the instinct to list all the reasons something will not work.
Decision Principles
Principle: Let customers tell the story
- When: a founder or product leader wants to announce a big new strategy.
- Why: Masad says the temptation is to make big news and declare a new direction, but the better path is to hold a general goal and move toward it incrementally. Each increment is an individual thing that comes up with a customer, and once you help one customer do it, everyone can.
Principle: Comply first, then improve
- When: an enterprise demands controls you think are wrong, such as per-person token budgets.
- Why: Masad says in the short term you have to do what they ask because "that's how you get in." Then you invent something better, such as a way to see where value is being created, and show them the way.
Principle: Shut down a dying business
- When: a fast-growing product line depends on a market that the product's own direction will remove.
- Why: Replit Bounties, where teenage hackers on Replit took paid coding jobs using early AI, grew fast. Masad shut it down because people would eventually ask agents for the entire thing directly, without going through the teenagers.
Trade-offs & Nuance
Guardrails versus speed and cost
Replit puts guardrails on agent-written code, which Masad says sometimes makes the product slower and more expensive. Enterprises value this. Consumers are hard to get to care about security, and Masad says some competitors do reckless things with insecure apps. The same choice that wins enterprise deals can lose on price or speed elsewhere.
Serving users versus serving IT
Graham asked whether Replit should give enterprise IT the rules it wants, or quietly help ambitious employees get around the bureaucrats. Masad called it "a tough one." Much of the enterprise work is giving admins and IT controls over what people can do and setting budgets, which Graham likened to sandboxing for humans.
Enterprise versus the original mission
Moving into enterprise created an identity question inside Replit, since enterprise did not feel like part of the mission. Masad resolved it by pointing out that Replit's enterprise users are entrepreneurial, ambitious people inside companies, often someone in sales building a tool for their own team. He says they tend to get promoted because of the impact they have.
Form factor matters as much
Masad thinks the moment that convinced the wider world that AI would write all the code, around the end of 2025, came from Claude Code and especially OpenClaw, even though Replit had run agents to completion for about a year and a half. He believes OpenClaw's form factor, including being able to text it, made the idea land for everyday people. His lesson is that he had assumed a good vision would seem obvious, but in reality you need to tinker a lot with how people experience it.
Common Mistakes
Mistake: Avoiding watching users struggle
Masad admits he did not like watching users, and felt physically ill when they were confused or hit a bug. Haya Odeh, his Replit cofounder, who forced him to do it early on, made being close to users a core part of Replit's culture. Graham's view is that it feels mortifying because until then you were in denial about how bad the experience really was.
Mistake: Holding on to your craft
Masad says many people are rigid about their identity and what they do. He has started over many times and is fine throwing work away when computers get good at something. He expects everyone to go through denial, depression and eventually acceptance about AI, and he says he went through that grief himself around 2020.
Mistake: Vague positioning claims
When Masad described Replit as the way AI flows into companies, Graham warned that he has to make that concrete. Anyone can say it, and someone at IBM could claim the same line. Masad adds that competitors copy what Replit popularized, such as "idea to app fast," and Graham notes they copy a frame from six months ago.
Practical Application
Watch users work in person
Invite a few users to sit with you and do a real task end to end, including any other tools they switch to. Look specifically for manual loops users run between your product and another tool, like Replit users pasting errors into ChatGPT. Those loops are candidates for automation.
Look for paid workarounds
List what your target users currently pay others to do, such as agencies, freelancers or outsourcing shops. Graham's lesson from Replit is that a market of bad, expensive workarounds is a sign of demand. Compare your price to that workaround, as Replit's roughly $1,000 compares with about $100,000 for an outsourced app.
Measure AI value per person
If your company rations AI tool budgets evenly, propose tracking which people and projects create the most value from that spend. Masad's observation is that ROI from AI tools follows a power law, so an even per-person budget starves the few people producing 100X results.
Write down and revisit predictions
Keep a short list of predictions about your market for the next one to three years, each with a date. Review it every quarter, note what you got wrong, and update it. Masad describes this as the main skill behind Replit's timing.
Map hidden depth to buyer needs
List the unglamorous infrastructure your product already has, such as reliability, security, permissions or reversibility. Then check which of those an enterprise buyer, as opposed to the end user, would pay for. Replit's years of systems work became its enterprise advantage.
Questions to Consider
- Which manual workaround do our users perform between our product and another tool that we could automate entirely?
- If the scaffolding we built around today's AI models becomes 10 times simpler as models improve, what part of our product would still be hard for a competitor to copy?
- Where in our product does the person who buys it differ from the person who uses it, and are we making the product worse for users to satisfy the buyer?
- If making our product 10% easier could double the number of users, as Paul Graham argues for consumer products, which piece of friction should we remove first?
- Is any fast-growing part of our business built on a market that our own product direction will eventually eliminate?
Bottom Line
Replit was early to coding agents because it watched users closely, had spent years on boring infrastructure, and had been predicting this future since its first pitch deck. In an AI market that commoditizes everything within months, the durable edge is constantly moving to the next increment while customers show you where to go.
Resources Mentioned
| Resource | Type | Why it was mentioned |
|---|---|---|
| The Self-Driving Company | Blog post | Replit's post describing its expanded vision of how firms will be organized when AI agents build and run much of their software. |
| Keep Your Identity Small | Essay | Masad suggests people who are rigid about their professional identity should read this Paul Graham essay. |
| Ray Kurzweil's writing | Author | Masad read Kurzweil early on, and credits his argument about exponential compute growth for how he predicts the future. |
Concepts to Explore
Running agents to completion
An agent that keeps working, fixing errors and retrying until a goal is met, as opposed to a model that answers once per turn. Masad says this was the key thing Replit demonstrated in 2024, and that the stopping condition should be that the program works.
Agent harness
The engineering around a model that turns it into an agent, such as edit tools, error detection and loops. Masad describes early problems the harness had to solve: models hallucinated line numbers, so Replit built its own text-replacement edit tool, and models got "lazy" and left to-do notes mid-function, which Replit had to detect and loop over. Today, he says, agents can do most of this with a plain shell.
Self-driving company
Replit's term for a future where AI agents increasingly build and operate the software a company needs. Replit treats itself as "customer zero" and a lab for figuring out how this changes work, including how to identify which people are doing valuable things.
Case Studies Mentioned
Replit Bounties shut down
Replit Bounties let teenage hackers on the platform take paid coding jobs, and they were already using Replit's early AI to do the work. The program grew fast. Masad shut it down because he concluded people would eventually ask agents to build the whole thing themselves. The lesson is to judge a product line by where the market is going, not only by its current growth.
$10 million project for $250,000
Masad says large companies, including banks, find a few employees inside who have adopted Replit, and those people start replacing expensive SaaS and old software. It usually starts with internal tools. One customer told Replit it had scoped a project at $10 million and instead spent $250,000 on Replit. Masad believes the outsourcing industry is largely finished as a result.
Telecom hackathon flipped sentiment
At a hackathon with a very large telecom company, employees were surveyed before and after using Replit for about an hour. Before, roughly 80% said they were anxious or worried about their jobs. After, roughly 80% were optimistic. Masad uses it to show how far most of the world is from the AI experience Silicon Valley already has.
Tools & Products
| Tool / Product | What it does | Why it was mentioned |
|---|---|---|
| Replit | A browser-based platform to write, run and deploy software, now with an agent that builds apps from plain English. | The subject of the episode, from its 2016-era pitch deck to its enterprise push. |
| ChatGPT | OpenAI's AI chat assistant. | Early Replit users pasted code and errors between ChatGPT and Replit, which inspired the agent. |
| Claude Code | Anthropic's coding agent. | Masad and Graham discuss it as part of what convinced the wider world that AI would write the code. |
| OpenClaw | An AI agent people can message, including by text. | Masad thinks its form factor was the first time many people experienced an AI running a task to completion. |
People to Follow
Amjad Masad
Co-founder and CEO of Replit. He has wanted to make programming accessible since 2008, when he wished he could program on his Symbian phone in class, and Replit's original deck around 2015 or 2016 predicted training a model on millions of people coding.
Paul Graham
Co-founder of Y Combinator and an early Replit investor, known for his essays on startups. In this episode he supplies the super-linear ease-of-use principle and pushes Masad to make Replit's vision concrete.
Andrej Karpathy
AI researcher who, according to Masad, called the September 2024 launch of Replit's agent a "feel the AGI" moment.
Experiments to Try
Experiment: Remove one step, measure adoption
- Hypothesis: following Paul Graham's super-linear ease-of-use principle, removing one meaningful step from onboarding will raise activation by much more than the step's share of effort.
- Change: pick the single most common point where new users stall and automate or remove it.
- Test: ship it to half of new signups and compare against the current flow.
- Measure: the share of new users who complete a first real task, not just signups.
Experiment: AI value tracking pilot
- Hypothesis: a few people produce most of the value from a team's AI tool spend, as Masad says Replit sees across customers.
- Change: for one month, have each person log what they built or automated with AI tools and estimate the time or cost saved.
- Test: rank people and projects by estimated value per dollar of AI spend.
- Measure: whether value is concentrated in a small group, and whether an even per-person budget would have limited them.
Notable Quotes
"you have to just accept that everything that you're doing is like getting commoditized constantly." (Amjad Masad)
"If you make it 10% easier, you don't just get 10% more, people get twice as many." (Paul Graham)
"Arguably, the whole existence of the outsourcing shops was evidence of a market there." (Paul Graham)
"I disagree with rationing tokens." (Amjad Masad)
"No, let the customers tell the story." (Amjad Masad)
