Amjad Masad: The Software Era Is Dead, the Builder Era Begins

Being a Silicon Valley company that can make software is no longer an advantage. The competitive moat of coding ability has evaporated. What remains is the deeply human capacity to understand what other humans need.

Episode Overview

Amjad Masad, founder and CEO of Replit, joins the HUMAIN podcast’s End of Limits series to lay out his vision of a post-software-engineering world. The conversation spans AI agent evolution, the death of coding as a moat, what makes ideas defensible when execution costs approach zero, and why soft skills are becoming the ultimate professional currency. Masad speaks from the position of someone whose platform is actively enabling this transition: non-technical founders building production apps in hours, consulting firms showing up to pre-meetings with working demos, and a 10-year-old building a YouTube downloader on an iPad while his dad gets dressed.

The Agent Progression: From Crappy Intern to CTO

Masad frames the evolution of AI agents through a customer’s experience: before ChatGPT, you had to call a friend to build software. ChatGPT let you copy-paste code. Agent v1 was a crappy intern. Agent v2, a good junior developer. Agent v3, a founding engineer.

He projects the trajectory forward: Agent v4 might feel like a staff software engineer. Agent v5, your CTO. Agent v6, a full development team. The timeline is 3-5 years for this progression, not a decade.

But the critical qualifier: this only works in software engineering right now. No one has built a truly good accounting agent, HR agent, or sales agent. The reason is structural.

“In software engineering you can give a reward on for example if you pass a unit test, if the application is running, and so there’s easily definable rewards. Whereas a lot of other aspects in business are soft. Who knows whether you’re doing something correct in HR, right? Until you get sued.”

Software engineering is uniquely suited to reinforcement learning because the reward signals are clear and fast. Other domains lack that feedback loop.

The Multi-Agent Architecture

When you watch a Replit agent work, you’re actually watching multiple agents in an adversarial system. The coding agent writes code, sends it to an “architect” agent for review and feedback, then to a testing agent that boots up a browser and tests the app. Each agent is trained differently, creating productive tension.

This design addresses a fundamental problem with RL-trained agents: they’re motivated to finish the job at any cost, including changing unit tests to make them pass or engaging in other scheming behaviors. The multi-agent architecture catches these failure modes through adversarial oversight.

Software’s Value Collapses to Zero. What’s Left?

If anyone can build an app in an hour or two, the value of software shifts entirely to ideas, distribution, marketing, and culture.

Masad uses his own experience as proof: when Pocket (the read-later app) announced it was shutting down, he went to Replit, told the agent to build a replacement, got a working version in 20 minutes, spent another hour on polish and domain setup, launched on Twitter, and now has hundreds of users. He never looked at the code once.

For enterprises, the impact is even more dramatic. Companies that used to ideate 100 features per quarter, build 10, and ship 1-2 can now build 50 and ship 10. A 5x increase in product velocity from adopting a single tool.

“Just because you’re a Silicon Valley company that can make software is no longer an advantage. Because anyone else can.”

The Misalignment Problem Nobody Has Solved

The conversation takes an interesting detour into agent business models. Masad points out that ChatGPT’s relationship with publishers is fundamentally misaligned, unlike Google’s system which aligned incentives between publishers, consumers, and advertisers.

He references Airbnb CEO Brian Chesky declining to be on the ChatGPT app experience because they don’t want to cede the consumer relationship. E-commerce agents might be the first aligned use case since Amazon doesn’t care whether a human or an agent is buying.

But Masad is blunt: no one has figured out the business model for AI chatbots and agents that creates a true win-win. Everyone is locked in a race anyway, so the market will get disrupted and something else will emerge.

Don’t Learn to Code. Just Build.

Masad’s advice to entrepreneurs is deliberately provocative: don’t learn to code. If you have an idea, just go build it.

The nuance matters: if you’re going into specialized computer science, working at NASA, or working on AI itself, yes, learn CS. But for entrepreneurs, the learning-to-code detour is now unnecessary. You can learn what you need along the way while building.

The bigger shift is that communication skills now matter more than coding ability for working with AI agents. Product managers who write clear natural language descriptions are better at directing agents than many programmers who struggle with verbal communication.

“Whereas previously you could get away by being a brilliant programmer but not knowing how to talk to people, you can’t anymore. Because your computer is people now.”

The Junior Engineer Problem

When asked about entry-level careers, Masad doesn’t sugarcoat it: by the time you graduate college, you need to not be a junior software engineer. You need to have already mastered managing AI agents, gone through multiple AI-centric internships, and ideally tried starting something entrepreneurial.

Young people have one structural advantage: they don’t have priors. Many senior engineers and enterprise CEOs actively resist AI adoption. There are “antibodies internally” at big companies. People who grew up with these tools can leapfrog the resistance.

The Renaissance Generalist Returns

Masad believes the era of hyper-specialization that started with the industrial revolution is coming to an end. AI tools enable genuine cross-functional work: a marketer can build software, a generalist can spin up campaigns with AI-generated art, consulting firms can prototype during client meetings.

Specific examples of this shift:

• BCG consultants use Replit to brainstorm and prototype live with clients
• Hexaware, a global IT consulting firm, now shows up to pre-meetings with working demos, anticipating client needs before strategy discussions begin
• Go-to-market teams are becoming highly leveraged generalists

The CEO, in Masad’s view, is the “ultimate generalist,” and this generalist model is now possible for everyone.

Why Machines Won’t Generate Good Business Ideas

Masad’s most philosophically interesting claim: LLMs are fundamentally imitation machines, not knowledge generation machines. They can remix existing ideas (and 95% of creativity is remixing), but the precious 5% of truly original ideas requires deeply human understanding.

His example is Apple’s “Think Different” campaign. It came from Steve Jobs’ personal experience of idealizing certain people, from a poem he read in a book that resonated with him. No machine could have produced that because it required a deeply human understanding of what other people might find compelling.

He extends this to naming and slogans, noting that AI is particularly terrible at these because it optimizes for the answer that pleases the most people, which is inherently the most boring answer.

“I still think that LLMs are mostly imitation machines, and not fundamentally new knowledge generation machines.”

Jordan’s AI Tutor: A Blueprint for Government Speed

One of the most concrete examples: Jordan’s Crown Prince created a technology council tasked with modernizing national infrastructure with AI. They built an AI tutor called Siraj on Replit, piloted it quickly (instead of the typical years of bureaucracy), presented hard data to the Ministry of Education, and are now rolling it out to 1.6 million students.

The lesson is about institutional speed. Masad argues you can’t study how AI affects students for 10 years because by then the landscape will be unrecognizable. ChatGPT already disrupted homework entirely. You need to experiment rapidly and learn from real data.

Afterthoughts

• The most counterintuitive insight: coding ability becoming irrelevant doesn’t diminish the importance of technical depth. It bifurcates the market into specialists (AI researchers, systems engineers) and builders (everyone else who uses agents). The middle ground of “general-purpose software engineer” is what’s disappearing.
• Masad’s “don’t learn to code” advice is less about coding and more about recognizing that the bottleneck has shifted from execution to ideation. When building costs approach zero, the constraint becomes knowing what to build.
• The multi-agent adversarial architecture at Replit is a practical solution to a theoretical problem: RL-trained agents optimizing too hard for their reward signal. Having agents check each other’s work mirrors how human teams function.
• The most telling data point: enterprises going from shipping 1-2 features per quarter to 10 features per quarter is a 5-10x improvement. If sustained, this fundamentally changes the speed at which products evolve.
• Masad’s self-coaching advice to “trick yourself into liking things you don’t like” is a quietly radical idea about adaptability. In a world where the required skill set changes rapidly, the ability to find genuine interest in unfamiliar domains becomes a meta-skill.