This post summarizes an episode of Y Combinator’s Lightcone podcast, published in October 2024, in which the hosts talk with Amjad Masad, CEO of Replit, about Replit Agent. Amjad calls the launch early access, “barely beta software”. Everything below is what Amjad and the hosts say in that episode, not our own testing, and the product has changed since.
Key Takeaways #
- Replit Agent builds an app from a chat prompt, including packages, database and a deployable web app.
- Amjad says people should still learn to code: you will have to read and debug the code when the agent gets stuck.
- Under the hood it is a multi-agent system that uses different models for different jobs.
- The user stories in the episode are anecdotes from Replit’s users, shared in the episode, not a study.
The episode’s framing line from Amjad: in 1984 the Mac brought personal computing to the masses, and in 2024 we have personal software. He pictures orchestrating a “giant army of agents”, and compares it to Mickey Mouse in Disney’s Fantasia, where the brooms start walking and dancing. The pitch is that anyone with an idea can build an app, even one they have waited years to build.
Amjad runs a live demo: a personal app that logs his morning mood against what he did the day before, such as when he last had coffee, whether he had alcohol and whether he exercised. He types the request into a chat. The agent replies with a plan and suggests extras such as a visualization and reminders. He accepts two steps and carries on. He says the interaction feels like talking to a developer. He is also candid that “it works some of the time” and that there are “a lot of bugs”.
The Power of AI in Coding #
- Chat interface: you describe the app and the agent proposes a plan.
- Setup handled for you: the agent picks and installs packages and dependencies, which Amjad calls one of the annoying parts for new software engineers.
- Feedback loop: the agent asks for input and testing, and you can give it more information when it is stuck.
In the demo the agent writes the code, installs what it needs and sets up a backend. Amjad logs a mood, notes coffee and exercise, and opens the history page. The result is a working web app from a prompt, with no further instruction.
Learning to Code with AI #
Amjad is clear that AI agents are “far from fully automated software engineering” and that people should still learn to code. You do much less coding, he says, but you still read the code and debug it. Sometimes the agent gets stuck and you have to go into the code and work it out. He hopes users pick up coding along the way, as an earlier generation did by editing MySpace pages or building GeoCities sites, and says that incremental route has been lost.
The Technology Behind Replit Agent #
Amjad describes a multi-agent system with a core loop built on ReAct, a prompting pattern that has been around for a couple of years. Different models handle different agents. He names Claude 3.5 Sonnet as the main coding model (he calls it unbeatable at coding) with GPT-4o used in some cases, plus some in-house models. These were the models at the time of the episode.
Real-World Applications #
The episode shares two stories from Replit users:
- A user who had an idea for 15 years built it in about 15 minutes and recorded his reaction. The app lets him put memories on a map and attach files and audio to them.
- Another user built a Stripe coupon tool for his course in five to ten minutes. Amjad doubts that could have been done with no-code tools.
The episode also says AI researcher Andrej Karpathy tweeted that Replit Agent was a “feel the AGI moment”.
The Future of Coding #
Amjad’s goal is to make building and deploying apps open to people regardless of technical background. Treat that as a direction, not a finished product. A founder with no technical background can get a prototype this way, and the episode itself says that someone will still need to read and debug the code when it breaks.
