Notes: How Replit Went From $10M to $100M ARR In Just 9 Months

This post summarizes Amjad Masad’s conversation with Y Combinator (July 2025) about Replit’s remarkable growth story, the technology behind AI agents, and his vision for the future of software creation.

The Origin Story and Mission

From Learning to Code to Replacing Code

• Replit started in 2016, went through YC in 2018
• Original mission: make programming more accessible
• After YC, became more ambitious: “a billion software developers”
• This sounded absurd at the time, but AI changed everything

Early AI Vision

• In the 2015 NLP hype, there was a glimpse of doing NLP on code
• Amjad had it in his seed deck: “at some point we’ll collect enough data to train models”
• GPT-2 in 2020 made it feel possible
• Built all the primitives: development environment, hosting, deployment

The Bet-the-Company Moment

The Pivot to AI Agents

• Tried agents in 2021 - didn’t work
• Tried again in 2022 - didn’t work
• Early 2024: “it’s getting close”
• GPT-4o could stay coherent for only 2-3 minutes
• The company was not doing well, burning too much money

The Layoff and All-In Decision

• Cut about 50 people, another 15-20 left
• Less than half the company remained
• Put everything on Replit Agents
• “Burn the boats” moment - this had to work

Claude 3.5 Saved the Company

• If Claude 3.5 hadn’t come out mid-build, they would have probably failed
• Claude 3.5 was the first model that could work for 5-10 minutes coherently
• The code generations actually worked

The Evolution of Agent Autonomy

Measuring Progress in Minutes

• Claude 3.5: 5-10 minutes of coherent work
• Claude 3.7: 45 minutes to an hour
• Claude 4.0/Opus: 7 hours (according to system card)
• 7 hours is basically a human worker’s day

The Limiting Factor: Computer Use

• Computer use “kind of sucks” right now
• This is the difference between one-prompt apps vs. babysitting
• Browser Use and Pig (Windows automation) are promising
• Weeks or single-digit months away from working really well

Advice for Founders

Take browser automation or Windows automation and apply it to enterprise verticals. The moment this technology works, those companies will explode.

Agent v3: Three Pillars of Autonomy

1. Transactional Infrastructure

• Built a fully transactional, reversible system
• Snapshot-based file system and database
• Every edit is an atomic snapshot in time
• Enables cheap copy-on-write forks
• Like Git made it safe for human programmers, this makes it safe for agents

2. Sampling and Branching

• Spawn multiple agents, each takes a shot at the problem
• Figure out which one works, choose that branch
• With sampling, SWEBench scores go from 70% to 80%
• Small models sampled can beat larger models
• “Sonnet sampled is probably better than Opus”

3. Computer Use for Verification

• True verifier is a test, not a judge model
• Computer use enables agents to do their own QA
• Sample out, pick the best branch that’s actually working
• Repeat for increasingly high reliability

Compute Budgets

• Give users the ability to set compute budgets
• Like O-style models: “here’s how much budget”
• “If you give us $1,000, we’ll spend them”

Who Uses Replit Agent?

The User Base

• People from every walk of life
• Product managers are a great use case
• PMs making significant business impact without talking to engineers
• Running A/B tests, optimizations independently

The Organizational Shift

• Created a new product group with designers, engineers, and PMs together
• All using AI all the time to prototype, sometimes going to production
• No waterfall model, no communication problems between teams
• Can move incredibly fast

The Bottleneck Shift

• Previously: infinite backlog, engineering time was the bottleneck
• Now: ideas get done, bottleneck is ability to have ideas
• “Looking at a to-do list that’s just empty - what do I do next?”

Security and Enterprise Challenges

LLM Security Issues

• LLMs are fallible, like humans
• Terrible at authentication - use old methods of salting and hashing
• Solo founders leaking API keys, bypassing login security
• No major catastrophe yet, but it’s coming

Replit’s Approach

• Built-in authentication component with captcha and security bells and whistles
• Integrated user management portal
• Partnered with Semgrep for security scanning
• Run code security scan before deployment

Components LLMs Shouldn’t Build

• Authentication (use built-in components)
• Payments (limited variations, use providers)
• Same pattern as humans: don’t write your own, use providers

The Technology Stack

Infrastructure is the Hard Part

• Distributed network file system took 2 years to build
• Nothing off-the-shelf for snapshot-based transactional systems
• Security is really difficult - crypto miners, abuse
• One of few places where you can get a cloud VM just by creating an account

NixOS Under the Hood

• Fully declarative, transactional operating system generator
• Multi-terabyte hard drives in all regions
• All packages cached and attached to every container
• Enables experimentation, rollback, sampling with agents

Moats and Compounding Advantage

• True moats often not obvious until decades into the company
• Not defensive moats, but compounding advantages
• Being ahead in some vector allows you to continue moving faster
• Infrastructure work is not as sexy as “we trained this model” but builds real lead

The Future of Coding Tools

The Spectrum

• One end: Power tools (Cursor, Windsurf) for developers to amplify efforts
• Other end: Consumer-facing “make an app” tools
• Replit in the middle: helping companies get stuff done for non-traditional developers

Visual Programming vs. Code Generation

• Visual programming never worked well - not Turing complete
• Code generation is Turing complete but interfaced via fuzzy natural language
• Synthesis coming: natural language interface with abstraction over code
• Like Smalltalk: interact with logical objects, not files

The Future of Work and Learning

Should Kids Learn to Code?

• For professional software developer route: CS degree and fundamentals still make sense
• For creators/generalists: not necessary to learn traditional coding
• Pick it up by osmosis - use Replit, run into issues, look at code/logs
• This is how our generation learned anyway

What to Teach Children

• Broad-based knowledge
• Be as generalist as possible
• Be as generative as possible - create lots of ideas
• “Once the making of things gets easier, the bottleneck goes back to how many ideas you can have”

Learn to Make Things

• Don’t put “learning to code” as the top thing
• Put “learn to make things”
• Learn to make things with code
• Learn to make things with video
• Learn to make anything with AI

The Future of SaaS

Current Reality

• Stories of people replacing hundreds of thousands of dollars of SaaS with Replit
• Example: Someone quoted $50K/year for org chart software, built it in 3 days
• HR professional at Replit built org chart software with ADP integration

The Prediction

• All application software will become dirt cheap
• No one will make money on traditional SaaS software
• Timeline: on the order of years, not tomorrow
• Today: 15% replaceable, Future: 100% replaceable

Key Insights

The Dystopian View is Wrong

• AI taking all our jobs is not correct
• The future of work is more human, more interactive, more multimodal
• Once making things gets easier, bottleneck shifts to ideas

For AI Startups

• Build on the very edge of what’s possible
• Start on a mission, technology not quite there yet
• Skate where the puck is going, it catches up with you
• Keep doing that - the technology will arrive

On Investor Confusion

• Investors use everything for 3 minutes, everything looks the same
• Products will start to diverge with different focuses
• Many investors are super confused about these systems