Netflix’s AI agent playbook: Stop fixing performance manually, build a pattern catalog

Netflix's Rajat Shah describes a system where AI coding agents ingest production profiling data to identify performance anti-patterns. The agent can pinpoint hot paths, search the codebase, and produce a fix with a code review in under 5 minutes. A real example found an O(n²) pattern consuming 8.8% CPU and proposed a fix with measurable

savings. The key innovation is a centralized pattern catalog stored as markdown files in a Git repo. This catalog serves as long-term memory for stateless agents, allowing them to recognize anti-patterns across services without re-learning. The catalog grows over time and can be bootstrapped from public resources like PyTorch's torchfix or internal playbooks. Shah stresses

that foundational automation—profiling, canary deployments, test coverage—must be rock solid before adding AI. The agent should not push fixes directly; it should run canary comparisons and only then open a code review for human approval. This reactive path then shifts left: the catalog can be used by reviewer agents or even during code authoring to

prevent inefficient code. The talk avoids hype by acknowledging the need for human judgment and proper guardrails. It presents a practical, stepwise approach: start with level 1 (manual but agent-assisted), then level 2 (orchestrated workflow with tools), and only later consider level 3 (full autonomy). The real benefit is catching performance issues earlier in the