AI EngineerVideoSource Linked

AI Agents for Performance: Ship Faster, Pay Less — Rajat Shah, Netflix

Netflix’s performance-agent playbook feeds production profiles and exact deployed code to agents, then relies on canaries and engineers to validate proposed optimizations.

AI Engineer · Jul 28, 2026

Open Source Open Markdown Open JSON

Source Summary

Netflix found a hidden **quadratic-time pattern** by giving an agent call-stack profiling data, then locating the matching code at the production commit. The proposed workflow turns profiles into candidate fixes instead of relying only on manual flame-graph inspection.

Practical Implication

Build a shared Git-backed catalog of observed performance patterns, and let coding or review agents consult it before inefficient code lands. A **stateful catalog plus a stateless LLM** can spread production lessons across services without requiring a vector database.

Agent-Ready Context

Netflix found a hidden **quadratic-time pattern** by giving an agent call-stack profiling data, then locating the matching code at the production commit. The proposed workflow turns profiles into candidate fixes instead of relying only on manual flame-graph inspection.

Build a shared Git-backed catalog of observed performance patterns, and let coding or review agents consult it before inefficient code lands. A **stateful catalog plus a stateless LLM** can spread production lessons across services without requiring a vector database.

Treat the profiler as an estimate and the **canary as ground truth**; error rates and an engineer’s decision remain release gates. The talk favors fixed, level-two orchestration because greater autonomy demands more evaluation, sandboxing, and prompt-injection defenses.

Context Map

agentcoding#harness-engineering#agent-memory#agent-reliability

Uncertainty

Treat the profiler as an estimate and the **canary as ground truth**; error rates and an engineer’s decision remain release gates. The talk favors fixed, level-two orchestration because greater autonomy demands more evaluation, sandboxing, and prompt-injection defenses.