AI performance engineering and tuning agents: AI Brendans or Virtual Brendans

Brendan Gregg's Blog

Brendan Gregg discusses the emergence of AI performance engineering agents built on his work, distinguishing between helper agents that interpret flame graphs and eBPF metrics versus "Virtual Brendan" systems trained on his publications. He explores challenges including pricing models, the limited scope of automation (only ~15% of actual performance engineering work), and the difficulty of commercializing such tools. The post covers Intel's $650M acquisition and subsequent shutdown of Granulate, an AI auto-tuner based on his work, and argues that while AI agents can help with previously-seen performance issues, they struggle with novel problems and quickly become outdated without continuous training.

On "AI Brendans" or "Virtual Brendans"