The Loop Is the Product — Roland Gavrilescu, Introspection

The first viral agent loop wasn't a coding agent. It was someone using OpenClaw to pit car dealers against each other for a better price. Roland Gavrilescu, co-founder and CEO of Introspection and formerly at xAI, lays out a blueprint for autoresearch in 2026 built on three ideas. The first is that the loop is the product: an agent's success depends on the quality of its signals and verifiers, and each loop's output feeds the next one. The second is that system distillation is the moat. Every loop's lessons, including evals, judges, skills and human judgment, should be captured as portable, versioned "agent recipes" that you own, independent of any model or provider. Introspection is releasing an early version called Pi recipes. The third is that valued work per watt is the score to optimize. Through a talent-sourcing agent example, he shows how to spot patterns in traces, calibrate judges with a human in the loop, and A/B test your taste with real users before promoting a change. Speaker info: Roland Gavrilescu X/Twitter: @rolandgvc (https://x.com/rolandgvc) LinkedIn: https://www.linkedin.com/in/roland-gavrilescu/ Related links: Introspection blog: https://www.introspection.dev/blog Timestamps: 0:00 Intro: from xAI to Introspection 1:03 A blueprint for autoresearch 1:13 Idea 1: The loop is the product 1:48 The first loop: haggling for a car with OpenClaw 3:03 OODA loops 3:38 Signals and verifiers 4:08 Looping the loop 4:23 Idea 2: System distillation is the moat 5:03 Recipes for AI systems 5:47 Agent recipes: reproducible frontier systems 6:53 Introspection and Pi recipes 9:17 Idea 3: Valued work per watt 10:47 Codifying taste into evals 12:12 Example: a talent-sourcing agent 13:17 Spotting patterns in traces 14:07 Calibrating judges with a human in the loop 15:12 A/B testing taste in production 16:47 Takeaways