Evidence receipt / prediction
Published · transcript-backedBret Taylor: prediction
10 Mar 2026 Cheeky Pint Bret Taylor of Sierra on AI agents, outcome-based pricing, and the OpenAI board
“" What would end up happening is a supervisor agent would observe your reasoning, say, "I think John should have actually looked up the policy here" and send it back with notes and say, "Actually, you're not allowed to make that decision.”
Source trail
Everything needed to verify it.
- Speaker
- Bret Taylor
- Attribution
- Verified speaker
- Claim type
- prediction
- Recorded
- 10 Mar 2026
- Publisher
- Cheeky Pint
Transcript context
…How do you force the LLMs? Mechanically, how do you force them to not answer off the top of their head, but actually look it up? We call it a constellation of models. Our platform, we call it Agent Studio, you essentially configure the goals and guardrails of a process. Goals and guardrails, not the sequence of steps because you want agency but you want guardrails around it. Within that, we'll use reasoning, but we use supervisor models to actually inspect that reasoning. If you were an AI agent in Sierra and you decided to go off-script, "I got this. " What would end up happening is a supervisor agent would observe your reasoning, say, "I think John should have actually looked up the policy here" and send it back with notes and say, "Actually, you're not allowed to make that decision. Here's the reasons why. Go redo that decision." It's a really effective technique. The way I think about it, which is a little simplistic, but I think basically right. If you imagine a reasoning system is right 90% of the time but has some either guardrail, malfunction, or hallucination 10% of the time, it's obviously better than that. Then you have a supervisor that's right 90% of the time. If you chain them together, you get 99% effectiveness. That methodology of layering reasoning and intelligence has been really effective. In general, it makes sense. You're basically layering, compute, you're layering reasoning on top of it. What's the neat about it, though, is we can abstract that complexity from our clients. They're expressing the goals and guardrails, and we have all these evals and tests and all these other things. We can find ways to make it more and more robust over time, but it doesn't require you to prompt engineer, write in all caps or whatever the hacks that people use to get these things to be conformant. You started in '22, '23?…
Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.