Notes: The Future of Software Creation with Replit CEO Amjad Masad

This post summarizes Amjad Masad’s talk at Y Combinator (September 2025) on the future of software creation and how AI agents are fundamentally changing the landscape.

The Evolution of Computing and Software

From Mainframes to PCs

• Mainframes were the first mainstream computing devices, requiring experts to operate
• PCs started as toys (Mac Paint era) until Excel made them useful for business
• Today, PCs (x86 computers) run the world economy, even in data centers

The Same Transition is Happening with Software

• Software engineering emerged in the 1970s with Unix and C
• Required 4-6 years of college education plus 2-3 years of on-the-job training
• Today, software is transitioning from something only experts do to something anyone can do

Replit’s Vision and Agent Strategy

The Mission

Replit has been working for almost 9 years to “solve programming” - making it so anyone can write software. They built:

• An IDE
• Language runtimes
• Online sandbox environments
• Deployments and cloud services

The AI Pivot

In late 2023/early 2024, Replit decided to put all resources into agents:

• At the time, agents barely worked
• But SWEBench (software engineering benchmark) showed clear progress
• From barely working in 2022 to 70-80% today
• The ultimate expression of their mission: making it so you don’t have to code

The Agent Habitat: Infrastructure is the Hard Part

Writing code is the easy part. The hard part is the infrastructure - what Amjad calls the “habitat” for agents.

What Agents Need

• Virtual machines - Ideally in the cloud, sandboxed (agents can mess up your computer)
• Scalability - Support millions of users
• Language support - Every programming language and package
• Standard Linux environment - Shell access, file read/write, package installation
• Deployments and databases - Everything a software engineer needs

Replit’s Built-in Solutions

• One-line authentication integration
• User management stored in databases
• Domain linking and deployment
• Secrets management for API keys
• Background jobs for continuous running
• Storage for grabbing things from the web

On the Roadmap

• Universal model access - Any model you ask for will be available directly in your app
• Payments - Not just for users to pay, but for agents to have wallets to pay for services
• Agent-to-agent integration - Your software engineering agent hiring accounting or sales agents

MCP Limitations

MCP is a more traditional RPC protocol - it doesn’t really solve the agent-to-agent problem. More interesting protocols are needed in this space.

Levels of Autonomy

Amjad draws an analogy to self-driving car autonomy levels:

Level	Description	Example
1	Language server	IntelliSense, lane assist
2	AI code completion	Copilot
3	First Replit Agent	Works with supervision
3.5	Agent v2	Works 10-15 minutes on its own
4	Agent v3 (current work)	Almost fully autonomous
5+	Future	Spin up 1000 agents with 95% reliability

Agent v3: Three Pillars

1. End-to-End Testing (Computer Use)

• Models can click around and use computers like humans
• Currently slow, expensive, and not very good
• But this is the frontier - will improve dramatically in 3-6 months
• Will enable agents to do their own QA instead of asking users

2. Sampling and Simulations

Replit built a fully transactional, reversible file system:

• Every edit is an atomic snapshot in time
• Enables cheap copy-on-write forks
• When facing tough problems, agents can fork themselves and the environment
• Try multiple approaches in parallel, find the best solution, merge it back
• Will increase agent reliability 2-3x

3. Test Generation

• Agents should generate tests for every feature they create
• Today, agents often break features they previously built
• Tests run on every change to prevent regressions
• Harder than it sounds - models are still pretty bad at generating unit tests

Predictions: Application Software Goes to Zero

The Core 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 or next year.

What This Looks Like

• Today: Businesses buy dozens of SaaS products
• Already: You can replace large parts with Replit Agent
• Future: 15% replaceable today → 100% replaceable

Real Example: Kelsey’s Org Chart Software

• HR professional at Replit, never written code
• Needed org chart software with specific features (ADP integration)
• Couldn’t find one that fit, and existing options cost tens of thousands per year
• Built it herself in 3 days using Replit Agent
• Now using it internally - could sell it as SaaS for tens of thousands per year

The Future of Work and Organizations

From Specialization to Generalization

Since the Industrial Revolution, the economy has trended toward specialization:

• One person makes one part
• Another tests it
• Another assembles it

But when your HR professional is also a software engineer, marketer, and anything else (because AI agents can do anything for them):

• Jobs become less specialized, less siloed
• Domain expertise is exponentially less important

The Generalist Employee

At Replit, they’re already restructuring:

• Product team = designers + engineers + product managers in the same person
• Merging roles to create generalist employees
• Org charts will look more like networks than hierarchies
• More like open source projects than traditional company structures

Everyone Becomes an Entrepreneur

• Employee mandate shifts from “write this marketing email” to “generate value for the business”
• Everyone is sort of an entrepreneur

The Sovereign Individual

Amjad references a book from the 1980s that predicted:

• Ideas will become wealth
• Merit wherever it arises will be rewarded as never before
• Anyone who thinks clearly will potentially be rich
• The information age (now “intelligence age”) will be the age of upward mobility

The Satoshi Example

• Single person created a trillion dollars worth of value
• Wrote the paper, wrote the software, put it out there
• We don’t even know who they are
• This will become a common occurrence

Universal Access to Opportunity

• Doesn’t matter if you’re in Silicon Valley or anywhere else
• If you can think clearly and use this technology, you can become a sovereign individual
• Access to opportunity will be universal

The Future of Collaboration and Startups

Assembling Teams at Light Speed

• Assemble groups of people quickly
• Assemble groups of agents quickly
• Create mission-purpose companies/projects and unwind them in a day or two
• Sometimes you might think you’re working with a human, but it’s actually an agent

Transaction Costs Go to Zero

Like getting an Uber today (one button):

• Getting a developer (agent or human) will be just one button
• Your agent can interview people/agents and find the best solution
• Build businesses at the speed of light
• Less reason to hire full-time employees

From Making Applications to Solving Problems

Today’s flow: Agent makes software → User uses software to solve problems

These are intermediate steps. Instead: Agents can just solve problems directly.

For Replit (and many businesses) to survive, they need to stop being focused on making applications and start being focused on solving problems with software.

Q&A Highlights

Will There Be Multiple Agents or One Universal Agent?

Multiple agents. People with unique domain expertise won’t share it openly - they’ll imbue it into specialized agents to scale themselves. There will be:

• Specialized domain agents
• Agents that assemble teams of agents
• Software development agents
• All potentially running through a main interface like ChatGPT

What’s Left for Humans?

When AI can automate most physical and cognitive tasks, what’s left? This fundamentally depends on your perspective on what makes human life meaningful beyond productivity.