A Developer in Your Pocket: Replit's Amjad Masad on What Happens When Everyone Can Build Software

Amjad Masad explains that Replit started from the observation that making software is hard largely because it is fragmented. A beginner has to download an editor, install a runtime such as Python or JavaScript, work out a package manager, and then figure out how to deploy and share the result. That cumbersome IT process, he says, is where many people get stuck and never learn to code.

Replit's answer is an end-to-end environment that is easy to learn, from the first line of code to deployment and monetization. Masad says the platform now has 34 million users worldwide, who are learning to code, building startups, and making personal and internal tools. A B2B package launched in July, and he says it is growing fast as people bring Replit to work.

Host Lenny Rachitsky recalls a viral story of an 11-year-old who built an app on Replit, and a skeptic who insisted she would still need hosting, a database and deployment. Masad says that is exactly what the platform abstracts away. The surprising part, in his view, is not the coding but "all the nonsense around it".

On the competitive landscape, Masad notes that Cursor is a VS Code fork with strong AI tools but is still an editor that needs a runtime and deployment environment. Some users even run Cursor alongside Replit. Being opinionated about everything makes Replit harder to adopt inside big companies' main engineering pipelines, but he says that is a trade-off the company accepts in order to serve product managers, designers, ops staff and even lawyers.

Masad starts the demo from Replit's home page, where users simply describe what they want. He pastes a prompt written by Replit product manager Aman Mathur: a Node.js web app where product managers can collect feature requests, let a community vote on them, show a status board with columns such as planned and in progress, and manage everything through admin controls. Masad himself often uses minimalist prompts and iterates, while PMs may prefer PRD-style detail, and he says both work.

The agent replies that it will build an initial prototype and suggests extras such as comment threads and email notifications. In the progress pane it creates a Postgres database, designs the schema and builds the home page. Masad says watching this is educational because users learn how web apps are structured, and he wants the process to be transparent even though many people never look at the code.

Because Replit already had real-time multiplayer coding, the team reused that system for the agent, which appears in the project as just another user. Masad contrasts this with chat-only tools such as v0: Replit is a full IDE where people can open the files, edit them or ask the AI to explain them. Rachitsky compares the experience to looking over an engineer's shoulder.

Asked about limits, Masad says the tool can get you an MVP and even some initial users. Trouble starts when you make large iterations, especially ones that change the app's structure and need database migrations. The agent is not good at those yet, and a bad migration can leave a non-coder stuck, though he says some users persist by asking ChatGPT or Claude for help. He expects improvement over the coming months.

Replit offers a Bounties feature for hiring human coders to finish a project. Masad would like the agent to go and grab a human by itself when it hits a wall, which he calls "sick" and Rachitsky notes is a reversal of roles. Meanwhile, the demo agent asks the user to confirm the app is running, catches an error, and starts fixing it proactively.

Masad estimates a decent engineer would need a few days to a week to build the same app, and a very good one perhaps hours. The agent took roughly five to ten minutes, and he guesses the compute cost around 15 cents. When the app runs, he tests it by submitting a feature and then asks the agent how to log in as admin, and it creates an admin account by running SQL queries rather than writing code.

Masad says most AI products aim to make engineers a bit better and capture a slice of that value. Replit asks a different question: what if everyone were a developer? Customers tell him they now do work they would otherwise have hired a developer for, and, because the activation energy is lower than going to a marketplace like Upwork, they also build ideas they would never have pursued.

He links this to the Jevons paradox: when something gets cheaper, total consumption rises. Cheaper electricity, for example, leads to more total spending on electricity, and he expects software to follow the same pattern. People will make far more software to improve their lives and work, and start more startups.

Small businesses use Replit for back-office tools, such as real estate agents managing their own data. Masad calls it a replacement for one-size-fits-all SaaS. In larger companies, product managers build a first version and test it with users before engineers put it on the roadmap. SpotHero's head of marketing built a competitive pricing tracker, and a partner engineer at X spins up prototypes to show customers how to use the X API.

Masad describes the stack in layers. At the bottom is a runtime that can install packages in any language, including native ones. Above it sit the editor and the infrastructure behind it, including the multiplayer system, and all of that is exposed to the AI. The agent can even take screenshots of the running app to check that the homepage renders.

He says a new discipline is emerging, which he calls AI-computer interfaces, as a counterpart to human-computer interaction. Making models use human interfaces, as in Anthropic's computer use, is expensive because of all the images and video. Replit instead gives the agent text representations of the shell, package installation and editor feedback. Masad says it is still more art than science, such as how many files a model can see before it hallucinates, but it is becoming more scientific.

On models, Masad says Anthropic's Sonnet is the best at coding and is the main one Replit uses, but the system is multi-agent and also uses OpenAI models, including manager, editor and critique roles. Replit trains some models itself, such as an embedding model for search. He says he wrote in 2022 that products would become a "society of models".

Looking ahead, Masad imagines agents that are tested by other agents and controlled through Slack, for example asking for Taylor Swift tickets the moment they go on sale. He argues that agents getting good at software is how AI becomes more general, since software runs our lives, the internet and businesses.

Masad says one persona he finds exciting is the startup CEO, such as Andrew Wilkinson of Tiny, who is a heavy user. These founders are often creatives who cannot code, and their ideas normally have to pass through someone else who must understand, find time and build them. With Replit they can build a v1 of their next product themselves, which engineers can then improve.

Inside companies, he says the pain of silos between designers, product managers and engineers comes from low-bandwidth communication through language, Slack and Zoom calls. The common denominator is code, so he asks what happens if the shared language becomes working prototypes. Replit's Figma extension, for instance, turns mocks into React code running on Replit, so engineers receive working code rather than screenshots.

Masad sees this as a radical reimagining of what most companies are, because everyone can be more general. He says Replit itself is fluid in this way: a designer there became a product manager, and design engineers sit on a spectrum between the two roles.

Asked which skills matter more, Masad points to being generative. He pictures a factory line where ideas are produced, built, and then consumed, and says the bottleneck used to be the middle stage, making things. With that opened up, you are limited by how fast you can come up with ideas, and he admits that despite being quite generative, he sometimes runs out of ideas now.

He also recommends learning some coding, but not the traditional way. He recalls that his co-founder, a designer, took a coding course whose first day was all about Git, giving her a tool before she had a problem. Non-engineers should not worry about tooling; they can learn by talking to an AI, debugging a little and fixing problems in something they built.

He then states what others have dubbed "Amjad's Law": the return on investment for learning to code doubles every six months. Even a little skill in prompting, reading code and debugging nets more power each time, because creating gets easier and the results more complete. Rachitsky reads this as engineering knowledge becoming more valuable even as agents write code, mainly to unblock the agent.

Masad thinks there is room for a business teaching AI-native coding, which would skip most of computer science and basic tooling and teach app structure, prompting and some debugging. He says Hacker News skepticism about AI coding tools makes sense for kernel work, but for building products the tools are already doing the building.

Rachitsky asks whether a Salesforce-sized business could be built entirely with tools like Replit. Masad says that if you believe AI is on a massive improvement curve, the answer is yes, though it feels absurd to his engineering mind. He cites Ray Kurzweil's point that humans struggle to grasp exponentials.

He explains how Replit applied that thinking. In 2022 the team built a Copilot-like autocomplete and trained its own models, but concluded that was not the right modality. The right one was chat plus an agent that creates things, and a year before the launch the models could not do it. The team built for the models expected in six months, and when they arrived, reasoning improved sharply. He sees roughly a six-month cadence.

Maintenance and debugging are already within the agent's reach, he says, but resilient architecture at scale, such as sharding databases and queue systems, will be the next bottleneck. Still, he can imagine a billion-dollar company with zero employees in five years. He adds an economic caveat: if software costs collapse, what can you charge, and can anyone build the next Salesforce? That is why he stresses idea generation and fast iteration.

Rachitsky offers an analogy to the "God of the gaps": humans are the gaps in today's agents, and those gaps will shrink. Masad agrees unless the current approach hits a fundamental limit, such as data, which he says he is not expert enough to rule out.

Masad advises founders and leaders to be resilient to fast change. Roadmaps are hard in anything AI can affect, he says, and when Anthropic released computer use, Replit threw out its plans and started building around it. He says being agile and willing to switch priorities immediately will be crucial.

He also urges flexible, low-structure cultures. At Replit people move along a scale from designer to engineer to product manager, and design engineers attend meetings on both sides. Rachitsky sums up the split: engineers become valuable by unblocking AI tools and debugging, while PMs and designers become valuable by discovering problems and articulating them clearly to the AI.

Masad says Replit was in open beta, with the agent available on its Core plan and an exit from beta expected within weeks. He previews a new product called Assistant, a more controllable cousin of Agent. Agent is like handing a developer a PRD, while Assistant is like sitting next to them and asking them to move a button three pixels. It responds in milliseconds to seconds, and is aimed at fast, reliable UI changes.