Replit CEO Amjad Masad on AI Agents, Company Automation and the Future of Apps

Replit’s bet on AI agents reaches beyond generating code: CEO Amjad Masad wants agents to help build and maintain software, route information inside the company, and eventually operate other apps for people. The company says its engineers nearly tripled their individual code output over six months while maintaining code quality. Masad’s account describes a workplace where people spend more time discussing ideas and directing agents, and less time typing code by hand.

That shift also challenges the business Replit is building. Masad predicts people will use fewer apps within three years, as agents take on tasks directly. Replit’s product helps users create applications, but his longer-term picture gives agents the ability to solve a problem without requiring a person to make or open an app first.

Those claims come from an interview with Platformer’s Casey Newton, and they describe the company’s own experience and expectations. They don’t establish that the same productivity gains or autonomous workflows will transfer to other organizations. Still, Replit offers a useful case study in how software companies are testing AI agents both as products and as internal tools.

From browser coding to an agent-led product

Masad co-founded Replit in 2016 with designer Haya Odeh. He grew up in Amman, Jordan, in a family with Palestinian roots, sold software for internet cafés as a teenager, and later worked as a founding engineer at Codecademy and on JavaScript infrastructure at Facebook.

Replit began as a browser-based place to run code without installing development tools. Its current product follows a different interaction pattern: a user describes an application, and an agent can design, write, test, deploy, and host it. The aim, Masad says, is to make creating software accessible to people who don’t identify as programmers.

He objects to the label “vibe coding” because it still centers coding. In his view, people want to make an app, a game, or another tool to solve a problem; requiring them to think of the activity as programming can put them off before they start. He sees code as the underlying material that lets AI systems carry out many tasks, not the main thing users should have to learn.

That distinction reflects Replit’s earlier purpose. Masad says the company wanted to erase the gap between learning to program and building something useful. His experience at Codecademy showed that coding lessons could help individuals change careers or create products, but learning syntax and working through technical details still appealed to a narrower audience. Replit’s premise was that users should be able to start by making things, with learning happening along the way.

Masad’s own view of coding education has changed sharply as AI systems improved. He started programming in Jordan as a child and was frustrated by the tedious parts of the work. When he helped start Replit, he expected natural-language interfaces to make programming easier over time. He recalls seeing GPT-2 in 2019 and thinking that language models might eventually translate everyday instructions into code.

In 2020, he wrote that people would still need to learn programming for the foreseeable future, but would have to learn less of it. The argument later became known as “Amjad’s Law”: the returns on learning to program were, he said, doubling month over month. That expectation held until newer models changed what he thought people needed to know.

By 2025, Masad says, the capabilities of models such as Claude Opus led him to conclude that most people no longer needed to learn to code to build software. He began making applications without inspecting their code, before the phrase “vibe coding” became widely used. For someone who had spent 15 years trying to make programming easier to learn, the conclusion was personally uncomfortable. He says he had to revise his beliefs to keep Replit aligned with what he saw as the technology’s direction.

Some engineers objected that people using AI to build apps could produce insecure software. Masad’s response is that AI can also help identify and address those problems, including through security agents. He says that at Replit, most engineers now inspect code close to deployment, and the company is working to automate more of the review process.

That approach changes where human attention sits in software development. People still need to define what should be built, assess whether it works, and decide whether it is ready to ship. But the typing and some review work can move to agents. How well this holds up depends on the quality of those systems, the risks of the software being built, and how effectively teams catch mistakes.

More building, less handwork

Replit says its engineers contributed 5.8 times as many lines of code in the first six months of the year as in the preceding period. The company also says code quality was preserved. That figure is a measure of output reported by Replit; the interview doesn’t give a detailed description of how the company assessed quality or whether the same increase appeared across every team.

Masad argues that higher code volume matters at moments when a company needs to find its next product direction. He describes software development as a field where success depends on trying many ideas. Replit’s first coding agent, he says, helped establish the company in a new phase of the market. He says it launched in September 2024, and that later coding agents from other companies followed. His point is that a company needs the ability to experiment quickly if it wants to find the next useful product.

For engineers, the practical change is less time with hands on keyboards and more time deciding what to ask agents to do. Masad describes Replit’s offices as more focused on discussion and whiteboard work than they were two years earlier. Engineers may send prompts through Slack, and an internal autonomous engineering agent can take a reported bug through to a proposed code change, or pull request. That can shorten the path from a customer request to a potential release.

Replit’s product ambition is to make that path part of one experience. Masad says he doesn’t want customers to switch between a collection of separate tools for design, coding, deployment, and growth. He describes Replit as one product that should take users from an initial idea through design and implementation, then into deployment and continued development.

The approach could help people try more ideas, but more generated code doesn’t automatically mean more useful software. Teams still need a way to check that a feature solves the right problem, works in its intended setting, and doesn’t create new issues. The interview emphasizes speed and experimentation; it gives less detail on how Replit measures the long-term cost of reviewing, maintaining, and securing a larger volume of software.

Design and technical work start to overlap

Replit is extending its agent model into design. The company has launched Replit Design, and Masad says much of the design work on the product’s Agent 4 release took place inside Replit itself. He argues that a working, interactive prototype can be more useful than a static mockup because people can experience how a design behaves.

His account draws on a familiar idea in product development: prototypes let teams test how something feels before committing to a finished implementation. Masad points to the original iPod as an example of a product whose physical prototype helped guide its development. He contrasts that kind of hands-on exploration with a process in which designers create fixed images and engineers later try to reproduce them in software.

At Replit, he says, design and engineering are becoming less distinct. The company has a Design Engineering group whose members work across both areas, and Masad says product managers can also take on parts of the design and engineering process. That doesn’t mean the underlying skills are interchangeable. He argues that design thinking remains valuable, even as the tools people use to apply it change.

Replit Design includes what Masad calls “ambient intelligence”: when a user selects an artifact they have created, the product can suggest variations. A user who doesn’t know the right design vocabulary can compare options, pick a direction, and ask the agent to identify the style. The idea is for the interface to help people make choices as well as execute instructions.

Masad also emphasizes the option to choose different models for design work. He says Kimi K3, from the Chinese lab Moonshot AI, is available in Replit Design and, in his experience, performs better as a designer than Claude. His explanation is that Claude tends to move quickly into code, while Kimi spends more time reasoning about the design problem. That is Masad’s assessment; the interview supplies no comparative benchmark or test results for the two systems.

The broader change is that AI tools can make it easier to move between a visual idea and a working product. A person can explore alternatives through an interactive prototype, then ask an agent to implement them. For teams, that may reduce handoffs between design and engineering. It also puts more weight on people’s judgment about which options are worth pursuing and whether the result fits the product.

A company that routes its own information

Masad’s “self-driving company” idea applies the same agent-first thinking to Replit’s internal operations. He says a large part of a chief executive’s job is routing information: connecting people who should talk, directing questions to the right team, and moving knowledge from one part of the company to another. Replit wants to automate some of that work so employees can spend more time on creative projects.

The first step is making company information easier to find. Replit uses an internal bot that employees can ask about the business and its metrics. Masad says a data layer connects sources such as the company’s Notion pages, its Postgres database, and its codebase. The bot can search across those sources and synthesize information into a report, rather than relying on one database or a person who knows where each answer is stored.

Masad sees that system as a way to make information flow less dependent on hierarchy. In his account, a company could begin to function more like a network, with employees able to ask questions and reach relevant material directly. The usefulness of such a system depends on whether the underlying information is current, accessible, and interpreted correctly. The interview describes the design, but doesn’t report accuracy figures or explain how access controls work across its connected sources.

Replit’s internal tools are also changing what it buys from other software companies. Masad says employees created tools that allowed Replit to cancel or reduce most of its analytics products. The company’s agent can write a Python program that joins data from its warehouse and production database, then present the results through an interface. Masad says that offers more flexibility than products that gather data through their own systems or connect to only one source.

He didn’t identify the specific vendor behind a previously discussed seven-figure contract, saying he didn’t want to name the company. Instead, he described a broader pattern: employees build an internal tool because they need something, and the tool sometimes works better for Replit than the product the company already pays for.

Masad says this work tends to emerge from employees’ own ideas rather than a formal search through vendor contracts. He gave the example of an engineer who had built a Git hosting service after experiencing outages with GitHub. The engineer showed Masad a demonstration, but Masad wasn’t sure whether it would become a product. He says he sees new projects at Replit each week; some might become customer-facing products, while others fill internal needs or replace paid software.

For Masad, the conditions that enable this experimentation matter as much as the tools. He says Replit gives employees room to build and tries to keep work from being intense all the time. The company works in release cycles, with two or three releases a year and periods of three or four weeks of concentrated effort. Between those pushes, Masad says, employees often have more room to explore ideas without a fixed deadline.

The ability to create and maintain internal applications could change the usual decision about whether to buy software or build it. Masad says agents can now help with the first version, then continue working on maintenance and iteration as users provide feedback. He imagines a lightweight process in which someone reviews requested features while an agent handles much of the ongoing development.

He expects this to weaken some traditional arguments for buying software, particularly the burden of maintaining a tool built in-house. He also says agents may become better at writing secure software and finding security problems. These are expectations, not guarantees: internal tools still need people who understand the data they handle, what failures would mean, and whether the agent’s changes are safe to use.

Growth, hiring and the effects on jobs

Replit’s rapid expansion gives context to Masad’s account of its internal changes. In March, the company raised $400 million at a $9 billion valuation, three times its valuation six months earlier. Replit says more than 50 million people build on its platform, including 85 percent of Fortune 500 companies, and that it is on track to reach $1 billion in annual recurring revenue by the end of the year. The source compares that target with $2.8 million in revenue for all of 2024.

These figures come from Replit’s own statements. They describe a company growing quickly, but don’t tell readers how revenue is distributed across customer groups or what costs it takes to serve them. The scale helps explain why Masad sees the company’s internal use of agents as a pressing operational question as well as a product experiment.

Masad expects AI to create more companies and more jobs in total, while also making companies smaller and prompting layoffs. He says the latter outcome is certain. His view combines a forecast of economic growth with an acknowledgment that individual workers may bear painful costs as organizations change.

He rejects hiring based mainly on elite credentials or narrow measures of achievement. Replit hires people he describes as ordinary, including young people who show energy, curiosity, and a willingness to learn AI tools. The company has interns as young as 16, he says, and also hires people in their late teens who are taking time away from college. He says a record of trying to build a company can matter more to him than formal recognition, even when the attempt didn’t succeed.

Masad says Replit expects to have more employees three years from now because he sees room for the company to expand into areas such as design and entrepreneurship. That expectation is specific to Replit’s plans. It sits alongside his broader prediction that companies may become smaller, which suggests he expects the number and type of businesses to change even as some expand.

He also recognizes the difficulty of telling experienced workers to replace skills they have used for decades. Masad calls that advice cruel, and says that while people should adapt and keep learning, the burden of transition can’t simply be put on individuals. His comments highlight a tension in the AI productivity story: new tools may create new work and new companies, while disrupting existing jobs and leaving people with uneven options for changing course.

Masad has been training models himself, including a cost estimator for Replit. He says one problem with agent-based coding products is that users can’t easily know whether a request will cost $10, $20, or consume the month’s available credits. His model was working better than he expected. That example points to a practical adoption issue: agents may make software creation easier, but uncertain costs can make it difficult for users to plan or set limits.

From apps to agents that act on intent

Masad agrees that many people may create their own software, but expects another shift to limit how many apps they use. Instead of asking someone to build an application as an intermediate step, he says, an agent could receive a goal and carry out the work directly.

He says he already uses Replit this way at times: rather than ask it to create an app, he asks it to do a task. Agents can also retain memories and build reusable skills that include software, application programming interface calls, and scripts. Those capabilities could help an agent handle repeated tasks without a person having to explain every step again.

Masad predicts that in three years people will use fewer apps and more agents, with agents using the remaining applications on their behalf. He also allows that as many as 50 percent of the apps people do use might be self-made. The forecast is not a measured estimate; it is his view of how the relationship between people, apps, and agents could change.

For that vision to work, agents must reliably operate software interfaces and handle tasks that may involve several steps. Masad predicts that agents’ ability to use computers will improve significantly over the next two to three months. He imagines people messaging an agent through a service such as iMessage or ChatGPT, then letting it use a browser or desktop software on their behalf.

He also sees a role for longer-running loops that can use existing software or create new software to solve a larger problem. The shift would be from issuing a series of detailed instructions to expressing an intended outcome, with the agent deciding which actions and tools are needed.

That raises a practical question about where the agent runs. Masad expects operating system makers to build stronger agents into their products, and argues that people should be able to ask a phone assistant to carry out a task such as booking a table. He thinks those systems will eventually improve, even though he believes some companies have been slow to act.

Replit’s position, he says, includes relationships with Microsoft and Google, while the company has had difficulty working with Apple’s app distribution system. He sees the larger opportunity in enterprise software, where there isn’t one operating system used by every organization. Replit could supply a kind of operating layer for companies by connecting their data and tools through agents.

This ambition links Replit’s product strategy to its internal automation work. The company wants customers to build software with agents, while also testing how agents can connect information and workflows inside a business. The enterprise pitch is less about making one more app and more about giving organizations a system that can coordinate work across the tools they already use.

Productivity is not the only measure

The interview closes on a question that complicates the promise of higher output: if agents generate more code, will people spend the saved time reviewing that code, producing another round of work rather than gaining time back? Masad says that if companies optimize only for productivity, there will always be pressure to do more.

He argues that teams should also ask whether the work is meaningful, enjoyable, and creative. For him, those goals could lead to more useful work while changing what people expect from technology. He questions whether economic output alone captures improvements in people’s lives, and says a meaningful working life can include income and security as well as creativity and good working relationships.

That answer is an aspiration, not a productivity metric. Replit’s reported increase in code output shows what the company says its agents have changed; it doesn’t show whether employees gained free time or whether the resulting work became more satisfying. The distinction matters for organizations adopting similar systems. Agents can reduce some manual steps, but the benefits depend on what companies choose to do with the capacity they gain.

Masad’s picture of Replit is therefore both a software product vision and an operating philosophy. People describe what they want; agents build, connect, and maintain the tools; employees spend more time choosing directions and less time on routine execution. His three-year prediction is that agents will become the primary interface to many tasks, with apps increasingly working behind the scenes.

Whether that happens will depend on reliable computer use, predictable costs, secure software, and systems that can handle feedback and maintenance over time. Replit’s experience suggests that agents can already change how a fast-growing software company works. The next test is whether those gains hold up when agents have to operate across the messy, interconnected systems businesses and their customers rely on.