The Future of AI Agents with Amjad Masad

Amjad Masad, founder and CEO of Replit, and Yohei Nakajima, Managing Partner at Untapped Capital, joined Village Global partner Ben Casnocha for a live masterclass with Village Global founders to discuss the future of AI agents. Takeaways from the conversation:

• There are two broad use cases for AI agents today: computer use (coding) and research. Research agents and computer-use agents are converging and getting ever more sophisticated. Complex tasks like travel booking are still hard for agents, but we may have Chief of Staff level agents by the end of 2025.
• Evidence suggests the number of minutes an AI can work and stay coherent is doubling approximately every seven months.
• There’s a significant “arbitrage” opportunity in quickly creating tools. A company was quoted $150,000 for a NetSuite extension, so an employee and Replit customer built the extension for $400 and sold it to his employer for $32,000. You can see the post on x.com.
• The lines between people who are “technical” and “non-technical” are blurring when it comes to using agents. People with strong systems thinking, the ability to break down problems, and grit, can build great applications. It’s valuable to be curious, fascinated, and interested in the systems. Amjad referenced the TED Talk on Grit by Angela Duckworth.
• There can be a U-curve where being too technical can make it harder to use the agent effectively. It’s counterproductive to force specific technical decisions rather than giving the agent freedom. Experts can sometimes be jaded or cynical, having seen previous AI cycles fail.
• It’s crucial to be comfortable with early first drafts and not giving up when the initial output isn’t good. Amjad wrote about perfectionism a decade ago. “Perfectionism is part excuse and part insecurity. There are traits associated with it that you need to debug and get rid of.”
• Technical capabilities are ever shifting, and companies must deal with the classic innovator’s dilemma. Companies have to “be bold and be willing to chew glass for a moment of time, see revenue go down, and see customers leave in order to sprint towards the future.”
• Moats are rarer than people think. “People will say they have a moat, but basically they’re saying they have a feature.” Amjad thinks Hamilton Helmer’s book, 7 Powers, offers the best understanding of moats
• Application companies are building their own models, but this is often due to external pressure or cargo-cult behavior. “If you’re not state of the art, no one will use it.” Niche models based on truly unique data or special reinforcement learning data can make sense for specific features.
• When evaluating early-stage founders building in AI, investors look for timeless principles like hard-charging, dynamic founders, technical skill, whether the domain is ripe for disruption, and if incumbents are slow-moving.
• Routine computer-based jobs like quality assurance, data entry, and entry-level SDR roles are within reach of automation, but the prediction of 10–20% unemployment in the next 1–5 years is high.
• We have yet to see agents generalizing outside of the training data but once they do, we’ve reached the singularity. “[The singularity] will be so disruptive that it’s not worth planning your business around it.”