Replit’s mission began with making programming more accessible, including tools for learning and setting up development environments. After years of agent experiments that fell short, the team made Replit Agent its central bet. Masad says Claude 3.5 was a turning point because it could stay on task for five to ten minutes.
How Replit Went From $10M to $100M ARR In Just 9 Months
Speakers
Replit founder and CEO Amjad Masad talks with Y Combinator hosts Dave and Tom about Replit Agent’s growth, the infrastructure behind autonomous coding, and how AI may change who builds software.
Masad describes transactional infrastructure that lets agents roll back changes and explore multiple branches. Replit snapshots files, databases, and virtual machines so teams can compare attempts and select a branch that passes a real test. He expects users to set compute budgets as parallel agent work becomes more useful.
Product managers and designers are using Replit to prototype, run experiments, and sometimes ship apps. Masad says this can shorten handoffs between product, design, and engineering, while raising questions about security, ownership, and on-call responsibility. Replit provides built-in authentication and runs Semgrep security scans, with agents able to try to fix findings.
Masad sees Replit serving a broad group of knowledge workers who want to solve problems with software. He describes evaluating models, building a snapshot-based network file system over two years, and using NixOS and cached packages to support cloud development environments. He also imagines interfaces that combine natural language with visual or conversational ways to understand what software is doing.
The hosts discuss Replit’s reported 45% compound monthly growth and the strain rapid adoption placed on its systems. Masad says the company tracks product and retention goals, while the hosts warn that revenue growth can hide churn and poor margins. They also discuss replacing some costly software and Masad’s advice to build at the edge of what current models can do.
Takeaways
- Treat model evaluation and user outcomes as ongoing product work, not a one-time model choice.
- Make agent actions reversible so teams can test multiple approaches and recover from mistakes.
- Give nontechnical builders secure defaults for sensitive features such as authentication and payments.
- When AI reduces implementation time, help teams decide who reviews, owns, and supports what gets shipped.
- Build early around a plausible model improvement, then make sure the product delivers value as capability grows.
Mentioned in this talk
Y CombinatorBrowser UsePIG
Anthropic
OpenAI
Semgrep
Google
NetflixThe BreakdownClaude 3.5
Replit AgentClaude OpusClaude Sonnet 3.7Authentication
Cursor
Windsurf Editor
Claude CodeAI Engineer mobile app
Granola
Fast ApplyAI agentsSmalltalk-80
NixOSTransactional infrastructureParallel agent samplingAgent evaluationSearch inside this talk
What was claimed 19
- ceo ofAmjad Masad CEO of
ReplitCEO and founder“The Future of Software Creation with Replit CEO Amjad Masad”
0:00inThe Future of Software Creation with Replit CEO Amjad MasadYouTube Verified4 sources· Talk
- foundedAmjad Masad foundedReplitFounder & CEO
“I'm the co-founder of Replit”
0:15inBuilding AI For AllYouTube Verified5 sources· Talk, Web· archived
- metricReplit: Participated in YC in 2018
“a YC company from 2018”
YouTube Unverified1 source· Talk
- metricReplit: Started in 2016
“Replit started in 2016”
YouTube Unverified1 source· Talk
- metricClaude 3.5: Could work for five to ten minutes and complete the job
“the first one that could work for five to ten minutes and actually do the job”
YouTube Unverified1 source· Talk
- metricClaude Opus: System card reported agents working for seven hours
“In the system card for Opus, they said they made it work for seven hours”
YouTube Unverified1 source· Talk
- featureBrowser Use: Works on browser automation
“working on browser automation”
YouTube Unverified1 source· Talk
- featurePIG: Works on Windows desktop automation
“doing the same for Windows desktops”
YouTube Unverified1 source· Talk
- featureReplit Agent: Supports rollback and transactional snapshots for agent changes
“The ability to roll back is very important”
YouTube Unverified1 source· Talk
- launchedReplit Agent: V1
“Replit Agent v1 to v2 was a huge jump in autonomy”
YouTube Unverified1 source· Talk
- featureReplit: Built-in authentication component with CAPTCHA and security protections
“we have built-in authentication”
YouTube Unverified1 source· Talk
- featureReplit: Runs a code security scan on deployment and gives a report of warnings and errors
“When you deploy a Replit app, we run a code security scan”
YouTube Unverified1 source· Talk
- integrates with
“a partnership with a great security company called Semgrep”
YouTube Verified2 sources· Talk
- featureReplit: Mobile app supports the vision of ambient building and agent interaction
“We have a really great mobile app”
YouTube Unverified1 source· Talk
- featureGranola: Records and transcribes meetings, with a searchable team record
“allows you to record meetings”
YouTube Unverified1 source· Talk
- metricReplit: Grew 45% compound monthly average since Replit Agent launched
“since Replit Agent launched, we're growing 45% compound monthly average”
YouTube Unverified1 source· Talk
- metricReplit: Distributed network file system took two years to build
“took us two years to build”
YouTube Unverified1 source· Talk
- uses
“Replit also uses NixOS under the hood”
YouTube Verified2 sources· Talk
- metricReplit: A customer replaced software quoted at $150,000 with an app costing $400 to make, then sold it to their employer for $32,000
“He went and made it in Replit and sold it to his employer for $32,000. It cost him $400.”
YouTube Unverified1 source· Talk
