Evidence receipt / belief
Published · transcript-backedBret Taylor: belief
10 Mar 2026 Cheeky Pint Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board
“On one hand, I think I actually agree with the principle that one of the powerful parts about AI, with its ability to do text, do audio, and I'll say, I'm not sure how you qualify computer use, but you can call it a form of image recognition and manipulation.”
Source trail
Everything needed to verify it.
- Speaker
- Bret Taylor
- Attribution
- Verified speaker
- Claim type
- belief
- Recorded
- 10 Mar 2026
- Publisher
- Cheeky Pint
Transcript context
…Do they switch to the agent language? They don't. We've done English over the publicly-switched telephone network. You have TCP/IP and English over PSTN. It reinforces, I guess, Dario's point, which is you can engineer all these fancy protocols, but the rails that are already there already exist. English is spoken by all AI agents, and the public switch telephone network has been around for 100 years, and it all works, which is fascinating. You have all these fancy MCP things, and we're doing English over PSTN. On one hand, I think I actually agree with the principle that one of the powerful parts about AI, with its ability to do text, do audio, and I'll say, I'm not sure how you qualify computer use, but you can call it a form of image recognition and manipulation. Certainly that's useful because you get to the point where you don't need to fully finish the last mile to get value. The thing I'd say, though, going back to talk about all the actions, and they're not all in the product, all the APIs don't exist. These visual interfaces were designed for us. Think of... Stripe is, I think, famously, it was one of the few enterprise software companies with good design for a long time. The Stripe dashboard is really elegant, right? And most enterprise software, you can't say that about their dashboard. I don't think the ideal agent harness will be that elegant because it's optimized for something else. It's optimized for the context that you need to perform complex multi-step procedures on behalf of a person. And my guess is it's just very different. I think seeing the way you write a harness for a software or a coding agent is just so different than the way you do UI design. I'm certain that it's great that you can click around a green screen or whatever—click around a green screen because it's an oxymoron—but type around a green screen or click around a legacy on-premises enterprise software system. I think these harnesses will be really good. I wonder if, is there a world two years from now where Stripe's ability to work with the agent that manages commerce for a direct consumer e-commerce company, that will be one of the ways you're evaluated. In fact, if your... For lack of a better word, harness is not compatible with the way their agents work, that's actually like you're not compatible with them. I'm not sure that's right. I think it's great that these things are backwards compatible. It's great that our agents have spoken over the telephone already in English. But I don't think it's the long-term future because there's so much value that you can provide. Put another way, the agents using a sophisticated application harness can just do a lot more and do a lot more with higher fidelity. Yes. Well, we should get to Sierra because you see a lot of real-world AI adoption. Maybe start by grounding us. What is Sierra? The business has scaled very quickly. One of the latest metrics that you can share because they keep changing from month to month as you guys grow.…
Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.