Amjad Masad on Replit, Coding Agents and the Future of Software

Replit co-founder and CEO Amjad Masad joins Harry Stebbings to discuss AI coding agents, software economics and the future of development. His company figures, product comparisons and predictions reflect his position at Replit and are not independently verified results.

Masad says technology took time to catch up with Replit’s original vision of making programming accessible. Growing up in Jordan, he experienced software as a way to create opportunities and start a business. After working at Codecademy, he wanted to make building software available to many more people.

He says Replit tackled browser-based development, hosting, package management and collaboration, but learning to code remained a barrier for many prospective builders. He now argues that some people can focus on creating products without learning traditional programming. That is his view of a changing set of tools, not a claim that programming knowledge has lost all value.

Masad attributes a major change to agents that can take actions over longer periods. He says Replit initially built substantial infrastructure and guardrails around models, then removed some code as newer models improved, before adding new scaffolding for more autonomous work. The release and performance descriptions are his account.

He describes model routing as a mix: Anthropic has been a core agent-loop provider for Replit, while Google models can handle some lower-cost tasks such as code search. He says the company has at times sent more tokens to Google, even while relying on Anthropic for the central loop. Usage varies with the task and time.

Masad calls Replit and similar companies “agent labs” that start with a user problem and select models or build infrastructure accordingly. He says Replit also develops some models itself. He is describing his company’s strategy, not an independent comparison of every agent product.

Asked whether building a proprietary model is a mistake, Masad says the answer changes as frontier and open models improve. He recalls Replit’s 2023 model outperforming GPT-3.5 on coding in its work, then says later frontier improvements narrowed the advantage. The comparison is his claim and no benchmark details are supplied.

Masad believes coding models may approach a plateau for some tasks, creating an opening for specialized models trained on relevant data. He says a lead lasting a few months can matter in an enterprise evaluation. The plateau, deal outcomes and the example of another company’s model are his assessment, not independently established results.

He puts performance ahead of cost while product capabilities are changing quickly. Once a task approaches a performance plateau, he says, a company can focus more on optimization and economics. His argument is conditional on the product and the pace of model improvement, both of which remain uncertain.

Stebbings asks whether $80 of every $100 in revenue goes to model providers. Masad says Replit’s share is significantly lower but still material, and says margins change with product investment. He does not give a current gross-margin figure; his claims about another provider’s margins are not verified in the interview.

Masad says Replit was close to profitability at one point before investing in a more parallel agent. He describes a cycle of building for performance and then optimizing the cost of the resulting product. His statements about agent speed and relative capability are company claims without test conditions or independent results.

He says evaluating models, selecting them for tasks, running proprietary benchmarks and A/B testing are core competencies for an agent company. Masad claims Replit can achieve better design results with Gemini than Google’s own products, while disclosing that Google is a partner. No comparative benchmark is provided.

Asked about Chinese models, Masad says he does not know their developers’ or government’s intentions. He distinguishes moral concerns from security considerations for enterprise data and says Replit had not yet taken that step, without ruling it out. His comments are personal judgments, not findings about any specific model.

Masad argues open models could support competition and reduce dependence on a small number of providers. He suggests a public-private effort to develop open models. His concerns about market concentration and pricing are a policy and economic view, not evidence of collusion.

Stebbings raises Jason Lemkin’s phrase “inference is the new sales and marketing.” Masad agrees that free tokens or relaxed limits can bring users into products, while recognizing that acquisition leaves a separate question of retention. He describes the appeal of building with agents as his own interpretation of user behavior.

Masad predicts product organizations may contain people with different technical, design and product strengths under a broader builder role, while engineers continue to work on infrastructure and other demanding areas. He also sees operations teams as potential users. These are forecasts and commercial observations from Replit’s CEO.

He says operations teams often combine software, spreadsheets and manual work, and describes customers building quoting or support workflows. He claims returns can be substantial but supplies no measured study or conditions. The examples do not establish that every organization will save money or reduce staffing.

On the so-called SaaS apocalypse, Masad says he is not seeing enterprises remove core systems such as Salesforce or Workday wholesale. He says some customers build on their APIs, while others use data warehouses as a foundation. His accounts are observations from Replit’s business, not a representative survey.

Masad agrees that alternative workflows could pressure some software growth and says certain survey or other point solutions may be replaced. He also expects new small competitors to undercut prices. These are predictions from a competitor with a commercial interest; the conversation does not quantify market-wide effects.

Asked about maintaining AI-generated software, Masad says Replit uses code review and optional testing workflows and is developing agents to monitor production systems. He describes security monitoring in enterprise deployments. The interview provides no independent effectiveness, coverage or safety measurements, and such tools do not establish that generated software is secure.

Masad estimates that 75% of Replit users are non-engineers and gives an illustrative customer-savings case. He argues some enterprise customers will pay for security and maintenance when they perceive a return, while consumers may be more price-sensitive. The user mix and return estimates are self-reported and lack a disclosed methodology.

He says Replit’s lower-priced core plan can function as a partially subsidized entry tier because inference costs make a fully free service difficult. He expects the company may subsidize it more as enterprise and pro revenue grow. The discussion does not disclose the plan’s actual unit economics.

On token prices, Masad distinguishes the price paid for a unit from the capability obtained for that price. He believes improvements have increased capability, while unit prices have not fallen as much as expected. More competition among models may change pricing, but he does not predict a specific reduction.

Asked whether Cursor is “dead,” Masad rejects a simple narrative. He argues the software-creation market may expand, some engineers want to inspect code in an IDE, and enterprise adoption can be sticky. His discussion of Cursor is a competitor’s opinion; the revenue number introduced by Stebbings is not verified.

Masad says he expects traditional IDE features to become less central as AI takes over some coding tasks, while acknowledging that people will still want to view and verify code. He specifically expects a need for development environments in safety-critical software. His “IDEs are dead” phrasing is a prediction, not a description of current universal practice.

For students, Masad advises against choosing computer science solely for an expected salary if they have no interest in the subject. He says foundations such as data structures and algorithms remain important, and sees paths into machine learning and other technical work. Whether university or self-study fits depends on the individual, he says.

Asked whether companies will shrink, Masad says he sees both leaner teams and customers who hire more as they grow. He cites an entrepreneur and a media company as examples. These are anecdotes from a vendor, not evidence of the overall employment effect of AI.

At Replit, Masad says he wants more builders and salespeople and fewer supporting roles. He sees sales increasingly involving education and helping customers change workflows. His own hiring preference and proposed small-business deal economics are not a general prediction of which jobs will disappear.

Masad says Replit’s app updates had been held in Apple review for about three months despite prior approvals. He says Apple cited guideline compliance and claims the company had responded. The interview contains his account, not Apple’s response or enough evidence to determine the reason for the review.

Stebbings speculates about Apple’s motives. Masad says he does not know and suggests the company may be considering how AI-created apps interact with its rules. He asserts Replit is not trying to bypass the App Store. These are his explanations and speculation, not verified findings about Apple’s decision.

Masad declines to name a competitor he wants to eliminate, instead arguing that Replit must keep innovating. He claims its agents helped start the current agentic movement and says rivals may copy its roadmap. Those are self-interested competitive claims rather than a neutral history of the field.

In the quick-fire round, Masad says he has changed his mind about sales hiring: although he values efficiency, he now wants to expand the sales organization faster to meet demand. He also wants a board member with experience building and taking a company public, alongside existing investors.

Masad names Elon Musk as an example of high performance because of the number of organizations he leads, while Stebbings names Roger Federer. Their examples are personal judgments. Masad describes being tired after leading one company and wanting to understand how others handle several.

Masad describes gradually changing his own habits over several years, starting with one workout a week and adding walks and other routines. He says sauna and cold exposure help him reset. This is his personal experience, not evidence that those practices or any specific regimen are appropriate for others.

Asked about parenting, he says he tries to put less pressure on himself after learning from experience with multiple children. This is a personal reflection rather than a universal parenting rule.

His closing lesson is that product-market fit can feel like demand pulling a product out of a founder’s hands faster than the company can supply it. He believes founders seeking venture-scale growth should continue testing and changing the product while keeping the broader vision. That is his retrospective heuristic, not a guarantee of success.