What I’m Learning About AI Agents at Y Combinator

Something fascinating is happening in AI right now, and I want to share what I’m seeing from inside Y Combinator’s Winter 2025 batch. Every Monday night, we have these incredible dinners where tech leaders come to speak candidly about where they think the industry is headed. Recently, we’ve had OpenAI’s Greg Brockman, Perplexity’s Aravind Srinivas, and Anthropic’s Chief Product Officer Mike Krieger all share their thoughts. Their message? 2025 is going to be the year of AI agents.

This isn’t just talk. YC just dropped their latest Request for Startups (RFS), and it’s packed with agent-related opportunities. But what’s really interesting is what we’re learning as we build in this space. Here are the key insights I’ve gathered so far:

1. We Desperately Need Better Dev Tools

Remember what it was like building web apps before good frameworks existed? That’s where we are with AI agents right now. Sure, we can build cool demos, but when it comes to production-ready systems? We’re still figuring it out.

We need better tools for everything: testing agent behavior (how do you even write tests for this stuff?), deploying agent systems (it’s way more complex than traditional apps), and monitoring what agents are actually doing. Right now it feels a bit like we’re building skyscrapers with hand tools.

2. Everyone’s “Shipping on Vibes”

This one’s actually pretty funny — and a bit concerning. In regular software development, you have clear metrics for success. Did the function return the right value? Did the page load in under 300ms? With agents, it’s more like “did it do what we wanted… kind of?”

Most teams (including us, if we’re being honest) are basically “shipping on vibes” rather than rigorous evaluation frameworks. We know this isn’t sustainable, but right now the tools and metrics for properly evaluating agent performance are still evolving. It’s both exciting and terrifying.

3. Vertical Agents are Where It’s At

One thing that’s becoming super clear: trying to build a general-purpose AI assistant that does everything is probably not the play. The real opportunity is in vertical-specific agents that do one thing really, really well.

Think of it like the early days of SaaS — the winners weren’t generic “business software” companies, but rather specialized tools for specific industries and functions. The same pattern is emerging with agents.

4. Yes, These Will Replace Some Jobs

Let’s be real about this one — AI agents are going to replace certain jobs. Not all jobs, and not immediately, but it’s happening. This isn’t about making people slightly more efficient anymore; it’s about full automation of specific professional roles.

The interesting part is how this is creating new types of jobs too. We’re seeing entirely new roles emerge around managing and directing teams of AI agents.

5. The Line Between Research and Products is Gone

This one keeps surprising me. The gap between “cutting-edge research paper” and “production feature” has basically disappeared. Things that were theoretical possibilities a few months ago are now part of actual products people are using.

It’s like if physicists were publishing papers about some new material, and companies were immediately using it to build buildings. The pace is unlike anything I’ve seen before in tech.

What This All Means

Being in YC right now feels like being in exactly the right place at the right time. The foundation for AI agents is here, but all the tools and best practices are still being figured out. It’s challenging, sometimes frustrating, but incredibly exciting.

The next few months are going to be crucial. The teams that can figure out reliable ways to build, deploy, and evaluate agents — while delivering real value to users — are going to win big. The technical challenges are significant, but the opportunity is even bigger.

Stay tuned. Things are about to get interesting.

Abhi Aiyer CTO https://mastra.ai/