Now Anyone Can Code: How AI Agents Can Build Your Whole App

Masad asks Replit Agent to create a mood tracker that records daily mood alongside coffee, alcohol, and exercise. The agent proposes a plan, selects Flask, vanilla JavaScript, and Postgres, then builds an app with a history view and activity trends. It also generates sample data and deploys the result, while asking Masad to do some quality checks.

Masad describes a multi-agent system using Claude Sonnet 3.5 for coding, GPT-4o for some tasks, and in-house tools for embeddings and code retrieval. The system combines text retrieval with lookups for code functions and symbols, and manages which memories enter the agent’s context. Masad says the agent can make mistakes or get stuck, so users may need to test, debug, or edit the code themselves.

Masad describes the agent as a coworker that can make errors and benefit from human guidance. He argues that learning to code helps people direct agents and solve problems the agent cannot handle. He compares incremental learning through personal software projects to earlier generations learning through Myspace pages and GeoCities sites.

Masad shares examples including a personal map for storing life memories and a Stripe coupon tool built in minutes. He says Replit Agent could help people move beyond the limits of no-code tools and eventually work with existing codebases. Planned directions include background work that returns a pull request, human experts joining when an agent gets stuck, and more user control through previews of proposed code changes.

Masad says Replit became more productive after flattening its organization and focusing on a few priorities. For the agent launch, teams across AI, the IDE, package management, and design formed an Agent Task Force. They reviewed live product runs twice a week to find failures, set priorities, and make rapid changes; Masad says reliability and support for more technology stacks are key next steps.