Summary
Rajat Shah, a staff software engineer at Netflix’s AI Platform organization, delivers a practitioner’s guide to using LLM agents for performance engineering at the AI Engineer World’s Fair 2026. The talk addresses a real tension in the vibe-coding era: AI coding agents ship code fast but often produce computationally inefficient implementations, because they lack context about platform-specific patterns and internal framework conventions.
Shah walks through Netflix’s experiment asking whether an LLM agent, fed raw profiling data, could identify performance hotspots as effectively as a human engineer — a task that typically takes 20+ minutes of manual flame-graph analysis. The answer was yes: the approach surfaces O(N²) implementations and other anti-patterns across polyglot codebases (Java, Python, Go), with identified fixes representing 0.5–4.6% CPU savings across affected services. The key insight is that profiler output is structurally similar across languages, making it highly LLM-readable.
The second half covers how to integrate this into the full software development lifecycle rather than treating it as a reactive production incident tool. Shah’s architectural recommendation centers on a stateful catalog of patterns and anti-patterns — stored as markdown files, not a vector database — that a stateless LLM agent can query, creating effective long-term fleet-wide memory. The talk is notable for its emphasis on practical implementation details over theory, making it directly applicable for engineering teams looking to reduce infrastructure costs as AI-generated code proliferates.
📺 Source: AI Engineer · Published July 28, 2026
🏷️ Format: Workflow Case Study
