--- schema: 1 type: video title: 'AI Agents for Performance: Ship Faster, Pay Less — Rajat Shah, Netflix' description: >- An inefficient quadratic-time pattern was hiding inside a tensor merge method on a live Netflix service, quietly wasting CPU on every request - nothing in a code review caught it, but it stood out clearly in a call stack. That's usually how this kind of waste gets missed - the normal process of catching it (profiling, reading a flamegraph, tracing it through the codebase, writing a fix, validating it) takes long enough that inefficiencies sit in production for months before anyone looks at them. AI-assisted coding adds to this, since it produces more code optimized for how fast it's written, not how fast it runs. Rajat Shah's team at Netflix built an agent to handle this loop: it took a call stack, traced the issue to the exact source line, proposed a fix, and validated it against a canary - a real production traffic split, with no regression allowed before it could ship. The result was based on measured outcomes, not the model's own assessment. The talk covers what it took to get an agent to a point where it could act in production, and walks through this specific case from start to finish - found, fixed, and confirmed by canary without anyone stepping in. It also covers what's next: the same pattern turned up independently in seven other services in one session, which is why it makes sense to build a shared catalog of patterns like this so they aren't rediscovered each time, and which could eventually be used to catch these issues in coding agents before the code ships at all. We'll also share the approach itself - how it was built, what it took to trust it, and what it caught - so other teams can adopt the same loop in their own SDLC rather than starting from scratch. Speaker info: - https://www.linkedin.com/in/rajatsshah/ - https://github.com/shahrajat - https://shahrajat.com/ Timestamps: 0:00 - Introduction: performance for ML serving 2:24 - The manual profiling loop today 4:46 - The experiment: can an LLM read a profile? 7:37 - From call stack to the exact method 9:07 - How the agent locates and reads the code 11:01 - First finding: an O(N) fix, canary confirmed 12:41 - The same antipattern across seven services 15:31 - Building a shared pattern catalog 17:16 - Storing and sharing findings across services 19:40 - Feeding the catalog to coding agents 20:56 - Human approval and verification 22:13 - Canary validation on real traffic 24:18 - Reactive vs proactive paths 26:54 - Catching waste before production 29:32 - Autonomy levels and what is next --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/CgsWxRUY5Eo-ai-agents-for-performance-ship-faster-pay-less 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/ai-agents-for-performance-ship-faster-pay-less-raja-89c835abf778 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/ai-agents-for-performance-ship-faster-pay-less-raja-89c835abf778.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T06:30:07.058Z' language: null children: [] videoId: CgsWxRUY5Eo durationSeconds: 2019 publishedAt: '2026-07-28T00:59:04.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 13 endSeconds: 147 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=13' heldReason: rights - ordinal: 1 text: null startSeconds: 158 endSeconds: 299 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=158' heldReason: rights - ordinal: 2 text: null startSeconds: 311 endSeconds: 457 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=311' heldReason: rights - ordinal: 3 text: null startSeconds: 311 endSeconds: 457 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=311' heldReason: rights - ordinal: 4 text: null startSeconds: 564 endSeconds: 669 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=564' heldReason: rights - ordinal: 5 text: null startSeconds: 680 endSeconds: 770 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=680' heldReason: rights - ordinal: 6 text: null startSeconds: 781 endSeconds: 935 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=781' heldReason: rights - ordinal: 7 text: null startSeconds: 935 endSeconds: 947 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=935' heldReason: rights - ordinal: 8 text: null startSeconds: 1038 endSeconds: 1050 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1038' heldReason: rights - ordinal: 9 text: null startSeconds: 1199 endSeconds: 1256 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1199' heldReason: rights - ordinal: 10 text: null startSeconds: 1266 endSeconds: 1333 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1266' heldReason: rights - ordinal: 11 text: null startSeconds: 1345 endSeconds: 1469 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1345' heldReason: rights - ordinal: 12 text: null startSeconds: 1480 endSeconds: 1624 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1480' heldReason: rights - ordinal: 13 text: null startSeconds: 1480 endSeconds: 1624 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1480' heldReason: rights - ordinal: 14 text: null startSeconds: 1636 endSeconds: 1778 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1636' heldReason: rights - ordinal: 15 text: null startSeconds: 1788 endSeconds: 2019 watch: 'https://liquidgold.cc/watch/CgsWxRUY5Eo?t=1788' heldReason: rights --- # AI Agents for Performance: Ship Faster, Pay Less — Rajat Shah, Netflix An inefficient quadratic-time pattern was hiding inside a tensor merge method on a live Netflix service, quietly wasting CPU on every request - nothing in a code review caught it, but it stood out clearly in a call stack. That's usually how this kind of waste gets missed - the normal process of catching it (profiling, reading a flamegraph, tracing it through the codebase, writing a fix, validating it) takes long enough that inefficiencies sit in production for months before anyone looks at them. AI-assisted coding adds to this, since it produces more code optimized for how fast it's written, not how fast it runs. Rajat Shah's team at Netflix built an agent to handle this loop: it took a call stack, traced the issue to the exact source line, proposed a fix, and validated it against a canary - a real production traffic split, with no regression allowed before it could ship. The result was based on measured outcomes, not the model's own assessment. The talk covers what it took to get an agent to a point where it could act in production, and walks through this specific case from start to finish - found, fixed, and confirmed by canary without anyone stepping in. It also covers what's next: the same pattern turned up independently in seven other services in one session, which is why it makes sense to build a shared catalog of patterns like this so they aren't rediscovered each time, and which could eventually be used to catch these issues in coding agents before the code ships at all. We'll also share the approach itself - how it was built, what it took to trust it, and what it caught - so other teams can adopt the same loop in their own SDLC rather than starting from scratch. Speaker info: - https://www.linkedin.com/in/rajatsshah/ - https://github.com/shahrajat - https://shahrajat.com/ Timestamps: 0:00 - Introduction: performance for ML serving 2:24 - The manual profiling loop today 4:46 - The experiment: can an LLM read a profile? 7:37 - From call stack to the exact method 9:07 - How the agent locates and reads the code 11:01 - First finding: an O(N) fix, canary confirmed 12:41 - The same antipattern across seven services 15:31 - Building a shared pattern catalog 17:16 - Storing and sharing findings across services 19:40 - Feeding the catalog to coding agents 20:56 - Human approval and verification 22:13 - Canary validation on real traffic 24:18 - Reactive vs proactive paths 26:54 - Catching waste before production 29:32 - Autonomy levels and what is next --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/CgsWxRUY5Eo-ai-agents-for-performance-ship-faster-pay-less 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/ai-agents-for-performance-ship-faster-pay-less-raja-89c835abf778 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.