ATRIUMsearch → argument graph
DataVideo · 25:48 — 26:35

Startups can now reach $15M ARR in 4 months with just a few people because AI agents allow for recursive self-improvement and skill loops.

Tan describes how building agentic systems involves identifying bottlenecks and creating software to remove them, leading to a recursive loop where companies can scale revenue rapidly with very few human employees. ✦ AI generated

Garry Tan · a16z Podcast · 2026-08-12 · original ↗

starts at this moment · 25:48

Elicited by

Talk about how YC is organizing around loops and how you're advising your companies to.

Building these agentic systems today, I think that's all you're doing is you're seeing where the bottlenecks are, and then, you know, almost every single time it's like, 'How do I just instruct the agent to create a piece of software or a markdown file that blows away that roadblock.' And then, when you take a step back and you've done that for the whole task, it's like, I mean, we we there are companies now today that, you know, we were we were talking about earlier that's like literally you can go from zero to 15 mil ARR in about 4 months with like two or three people and like hundreds of agent like, you know, a few hundred skill files.

verbatim transcript · starts at 25:48

Transcript · around this moment

25:42like uh, building these agentic systems today, I think that's all you're doing is you're seeing where the bottlenecks are, and then, you know, almost every single time it's like, "How do I just instruct the agent to create a piece of software or a markdown file that blows away that roadblock." And then, when you take a step back and you've done that for the whole task, it's like, I mean, we we there are

26:08companies now today that, you know, we were we were talking about earlier that's like literally you can go from zero to 15 mil ARR in about 4 months with like two or three people and like hundreds of agent like, you know, a few hundred skill files. >> It's amazing. So, there's this like recursive self-improvement and there's this hill climbing that happens. Like at what point does it start to fall

26:30over? Is there like do you need error correction? Do the errors compound? Does it get constrained by exception handling, human relationships? Cuz otherwise, these companies should just be able to grow and run forever. >> Yeah. I mean, I guess my experience once you get a lot of markdown files or yeah, anytime you have like some large amount of data, um, that's where like provenance matters. That's where um conflict management sort

26:56of you know when you have two facts if you have provenance then you're like well this is the more recent one like this is the one that wins. You need a cron job that's sort of sweeping through that and making sure that things are accurate and from the right places. Um but these are things that I think as you use the systems um you just know that you need to build

Around this claim