The Future of Software Creation with Replit CEO Amjad Masad

Masad compares today’s shift in software creation to the move from expert-run mainframes to PCs that became everyday tools. He says Replit spent years building an online development environment and, when AI advanced, turned its focus to agents so people could make software without coding.

Masad argues that writing code is only part of the challenge: agents also need a safe, scalable environment with access to Linux tools, packages, databases, deployments, and services such as authentication. Replit Agent V3 is intended to improve reliability with app testing, parallel trials in reversible environments, and tests for new features. He says these tools could let agents work for longer without human feedback.

Masad predicts that application software could become extremely cheap over the coming years, as people make tools tailored to their needs. He describes a Replit HR colleague who built an org chart tool in three days after she could not find a suitable product. He expects work to become less specialized, with employees taking on broader roles and company structures becoming more networked.

Masad envisions individuals using agents to turn ideas into products and groups of people and agents forming around specific projects. He says Replit will need to evolve from helping people make applications toward solving problems with software. In the Q&A, he says multiple agents will likely keep separate expertise and context, and that people with domain knowledge should consider building agents in their field.

Asked how agents can work autonomously for longer, Masad points to the environment around the model: Replit records changes as checkpoints so developers can inspect or restore earlier states. He says longer tasks need testing to limit goal drift, while shorter tasks benefit from reasoning and parallel attempts. He also predicts future software agents will learn through reinforcement learning and self-play, rather than relying only on human-written code.