ATRIUMsearch → argument graph
MechanismVideo · 7:54 — 9:24

Agent behavior splits into three distinct modes — looking something up, doing something (which raises credential/impersonation issues), and analyzing (which is powerful but where hallucination becomes a serious risk requiring step-by-step verification).

Stephen Sinofsky argues that the useful way to think about agents versus APIs is by what the agent is actually doing: lookup, action, or analysis — with analysis being the most agent-suited but most hallucination-prone. ✦ AI generated

Stephen Sinofsky · a16z Podcast · 2026-07-07 · original ↗

Share card ↓Export .mdExport .txt

starts at this moment · 7:54

Elicited by

“Stephen, do you have anything to add here just on the kind of like definitional territory that we're covering right now or kind of this discussion?”

What is the agent itself doing? Is it looking something up? Because that's actually a pretty lightweight thing that all systems are pretty good at... Then there's like I want to do something and that's where you get into very interesting issues... is it another paid seat, is it the same paid seat... and then there's the third thing which is analyze... it's also where hallucination really is a huge issue because you actually need a way to verify that every step of that analysis was correct.

verbatim transcript · starts at 7:54

Transcript · around this moment

7:54What is the agent itself doing? Is it looking something up? Because that's actually a pretty lightweight thing that all systems are pretty good at. And in fact, many many of the newlyannounced, you know, headless agent APIs are are just lookup and they're just you basically have a new interface to the old way to look something up, which is a lot more forgiving, a lot less UI goo

8:18and stuff like that. Then there's like I want to do something and that's where you get into very interesting issues over like well if you do something you have to be impersonating a specific person you have to have their credentials like it's a very is it another paid seat is it the same paid seat you have all these interesting enterprise software issues that come up if you actually want to cause a change

8:44to a system of record and then there's the third thing which is analyze. And so analyze is more than look something up. It's actually look up a bunch of stuff. It often involves multiple systems. And that seems very very tuned to an agent because you you're not time bounded. You can spend energy iterate. You can route it to different models and get different answers back and compare them. But it's

9:10also where hallucination really is a huge issue because if you're going to go and analyze something, you actually need a way to verify that everything every step of that analysis was correct. And so I think it's super interesting and important when you look at headless and agent which are conflated in you sort of have to figure out what you're talking about because we're on different places in the evolution the learning curve and

9:38the deployment of agents relative to sort of that three-way matrix >> matrix. >> Yeah. I I think this is actually a good leadup into a follow-up question, which is kind of like historically, what has made software sticky and how are agents starting to disrupt that? And I I leave that to either of you to answer. Maybe you guys can both kind of debate about that. Yeah, I'd say there there are many

✳ agent✳ API
Related moments