Netflix Engineer: AI Agents for Performance

Rajat Shah, a Staff Software Engineer at Netflix, shared insights on how integrating AI agents into performance engineering workflows can lead to faster shipping cycles and reduced infrastructure costs. Speaking at the AI Engineer World's Fair 2026, Shah presented a practitioner's guide, outlining a playbook for organizations looking to replicate these performance improvements.

The Problem: Performance Engineering Bottlenecks

Shah highlighted that while AI agents are accelerating code authoring, they don't always produce the most performant code. This is because these agents lack specific knowledge of platform-specific performance patterns and internal codebase nuances. The manual process of performance engineering, which involves triggering profiling, downloading raw data, analyzing flame graphs, and searching for code paths, is time-consuming and often only performed reactively when issues arise, typically at 2 a.m.

The Solution: Leveraging AI Agents for Performance

The core hypothesis is that AI agents can read profiling data, identify performance bottlenecks, and even suggest fixes much faster than humans. The experiment at Netflix demonstrated that by feeding structured profiling data to AI agents trained on common performance patterns (like O(N^2) loops, loop invariants, and inefficient memory allocation), these agents can accurately pinpoint issues.

The process involves several steps: parsing profiler output, identifying the exact code version running, filtering to relevant code, tracing the full call path, and then analyzing the data. Shah emphasized that the AI reads structured call stack data, not just a visual flame graph, enabling it to understand the context and identify problematic patterns.

Real-World Findings and Savings

The presentation showcased a specific finding where an AI agent identified an O(N^2) ImmutableMap accumulation issue, which was consuming 8.8% of total CPU time. The agent was able to draft a pull request (PR) for an engineer review within minutes, identifying potential savings of 6.5% in EC2 CPU usage and 3.0% in p90 latency. This was validated through canary deployments.

A second key finding demonstrated the power of AI in identifying recurring issues. The same bug, related to a counter object created on every hot-path call, was found in seven different services. Without a centralized catalog, each team had to re-derive the problem. With AI, the pattern was identified once, and then applied across multiple services, potentially saving significant CPU cycles (0.5% to 4.6%) across the board.

Integrating AI into the Development Cycle

Shah stressed the importance of building a long-term memory for AI agents, suggesting a centralized, ever-growing catalog of patterns and anti-patterns. This catalog, initially built from production findings and potentially augmented with public knowledge sources, acts as a foundation for future agents. The catalog can be simple, starting with markdown files in a Git repository.

The integration into the development cycle can occur in two ways:

• Reactive Path: Profiling, analyzing, fixing, and creating PRs, typically taking under 5 minutes and can be done on-demand or scheduled weekly.
• Proactive Path: Catalog lookup during PR creation, leading to inline comments and suggestions from the AI agent, preventing suboptimal code from reaching production in the first place.

Both paths share the pattern catalog, ensuring consistency and continuous improvement.

Guardrails and Best Practices

Shah emphasized that AI should handle the non-deterministic aspects, while deterministic steps should belong to the infrastructure. This includes ensuring robust test coverage, automating canary deployments for validation, and maintaining a human-reviewable pattern catalog. The final decision for code changes should always rest with an engineer.

The key takeaways were to build core foundations first, invest in automation, and start with the reactive path to gradually move towards a proactive, AI-assisted development cycle. The ultimate goal is to catch performance issues earlier, ideally before they are introduced into production.