--- schema: 1 type: video title: 'The Loop Is the Product — Roland Gavrilescu, Introspection' description: >- 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 --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/7taOQBfjDyE-youtube-a4f308b698332717ae75 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/the-loop-is-the-product-roland-gavrilescu-introspec-ebae3c242880 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/the-loop-is-the-product-roland-gavrilescu-introspec-ebae3c242880.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:45.106Z' language: null children: [] videoId: 7taOQBfjDyE durationSeconds: 1124 publishedAt: '2026-09-26T14:00:25.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 12 endSeconds: 68 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=12' heldReason: rights - ordinal: 1 text: null startSeconds: 92 endSeconds: 115 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=92' heldReason: rights - ordinal: 2 text: null startSeconds: 126 endSeconds: 183 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=126' heldReason: rights - ordinal: 3 text: null startSeconds: 195 endSeconds: 228 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=195' heldReason: rights - ordinal: 4 text: null startSeconds: 228 endSeconds: 239 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=228' heldReason: rights - ordinal: 5 text: null startSeconds: 249 endSeconds: 260 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=249' heldReason: rights - ordinal: 6 text: null startSeconds: 285 endSeconds: 309 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=285' heldReason: rights - ordinal: 7 text: null startSeconds: 320 endSeconds: 354 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=320' heldReason: rights - ordinal: 8 text: null startSeconds: 364 endSeconds: 422 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=364' heldReason: rights - ordinal: 9 text: null startSeconds: 433 endSeconds: 558 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=433' heldReason: rights - ordinal: 10 text: null startSeconds: 558 endSeconds: 568 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=558' heldReason: rights - ordinal: 11 text: null startSeconds: 568 endSeconds: 647 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=568' heldReason: rights - ordinal: 12 text: null startSeconds: 659 endSeconds: 740 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=659' heldReason: rights - ordinal: 13 text: null startSeconds: 817 endSeconds: 850 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=817' heldReason: rights - ordinal: 14 text: null startSeconds: 860 endSeconds: 913 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=860' heldReason: rights - ordinal: 15 text: null startSeconds: 923 endSeconds: 1011 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=923' heldReason: rights - ordinal: 16 text: null startSeconds: 1022 endSeconds: 1124 watch: 'https://liquidgold.cc/watch/7taOQBfjDyE?t=1022' heldReason: rights --- # 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 --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/7taOQBfjDyE-youtube-a4f308b698332717ae75 Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/the-loop-is-the-product-roland-gavrilescu-introspec-ebae3c242880 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:55) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:55 to 3:03) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:03 to 4:20) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:20 to 5:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (5:54 to 7:02) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (7:02 to 9:18) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (9:18 to 10:59) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (10:59 to 12:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (12:30 to 14:10) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (14:10 to 15:23) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (15:23 to 16:51) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (16:51 to 18:44) > Held: transcript rights do not permit this chunk's text.