Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board

**John** (0:01)
Bret Taylor is the ultimate Silicon Valley veteran. He was one of the creators of Google Maps, invented the like button, was co-CEO of Salesforce. He pushed through Elon's acquisition of Twitter, when he was on the Twitter board. He's now the chairman of the OpenAI board. And his day job is founder and CEO of Sierra, which is bringing AI to customer service. He's one of the smartest people I know, and the topic of how AI is changing established companies. Cheers. So, most important question, have you installed OpenClaw on your work laptop?

**Bret Taylor** (0:30)
I have not.

**John** (0:31)
Have you played with OpenClaw?

**Bret Taylor** (0:33)
I have played with OpenClaw. I haven't like bought a Mac Mini. I've, you know, you can put these things in virtual sandboxes for last money. It's really interesting. I mean, it's very compelling. It's probably the first, I wouldn't have predicted the first kind of broad, I don't know if consumer is exactly accurate, but maybe a hobbyist use of AI would have been this kind of semi-rogue open source project that goes through three name changes in three days.

**John** (1:00)
Yes.

**Bret Taylor** (1:01)
And I love it. I love everything about the chaos of it, just because people in our circles have been talking about AI agents for consumer use and all these fancy computer using agents. And instead, you're chatting over WhatsApp with a thing on a Mac Mini that is mildly unhinged and insecure. It's just fascinating. The whole thing is fascinating.

**John** (1:24)
But isn't that, okay, the thing that seems to me is funny, is if you look at the landscape, still in 2026, if you open a new Gemini chat, or if you open a new chat, it's basically a blank slate. There's no memory. And then, I mean, people talk about the WhatsApp and Telegram integrations and things like that. But it feels to me a big part of the value is not only can it do stuff proactively, but it has memory. But the way it has memory is this like super janky. It's like the movie Memento. It writes things to a markdown file, and it's just writing the things to remember. And the compaction is kind of buggy. It doesn't always write down the exact right things to remember and stuff like that. But isn't it funny that you can get super polished mainstream consumer apps that have no memory at all, or this wildly insecure three name changes project that kind of almost remembers things by scribbling notes in the margin, and that is the state of consumer AI?

**Bret Taylor** (2:21)
I have a probably not very thoughtful but kind of technical theory on this. So coding agents have gone through transformation over the past four months. Like the difference between if we're here in October versus now our conversation about the future of software engineering would be materially different and how often can you say that about a technology? And people always, in my circles anyway, you look at a coding engine and you extrapolate to other domains. You're like, could all digital tasks be like this? And the answer is obviously yes over some period of time. But it's really interesting because I think sometimes like, I think the hard part of engineering is in the details and code repos have very specific qualities. One is all the context is in one place, in files that are largely textual, not binary. And for most broad information tasks, that's not true. When you're writing your annual letter, my guess is the sources of information were in so many different systems, data warehouses. And so it's not like impossible for an agent to use those things, but the idea that you can straight line from coding agents to writing the Stripe annual letter, I don't totally buy.
And then similarly, when the agents actually perform work on a code base, there's feedback. There's compiler errors. There's often unit tests. There's integration tests. There's the history of every change every made in a really formal format, along with code reviews. And so you can actually, it's almost designed for a robot, and you can self-reflect. Maybe we as engineers are sort of like, have always modeled ourselves after robots, and now we can actually fully realize that vision, and so what's interesting about it is, like the idea that it wrote a markdown file for memory, I think is maybe more significant than a hack. I actually think to some degree-

**John** (4:14)
Turning your life into code kind of? Yeah.

**Bret Taylor** (4:16)
It's like you almost want to put all everything in a file system, that sort of looks like source code, not because that's the only way these agents can work, but actually it's quite an efficient way to get a mix of context and random access memory. If you think about like a vector database, it's more random access. You have to know what to look for, but actually that's not how real memory works. There's a mix of it, so you're loading a markdown file, and as you said, compaction, all these things matter. But the messiness of it actually probably produces a more useful agent than a lot of the fancier things. And I use memory in chat to be true, I love it. But I actually think this idea that there's a directory of just everything you've ever done is actually maybe more useful to an AI than people think. And actually, if you follow over the past couple of months, just this emergence of harness engineering, where you're building the harness around an agent to do work, I wonder if in the short term, it might just be one of those idiosyncrasies of history, like mimicking a code base is actually the best way to make a general purpose agent work.