AGENTIC AI FRAMEWORK FOR CONTINUOUS PERFORMANCE ENGINEERING

Abstract

We designed a multi-agent framework with four specialized agent types coordinated through a shared performance knowledge base. The monitoring agent continuously observes performance metrics and detects anomalies. The experimentation agent designs and executes performance experiments to characterize system behaviour and validate hypotheses. The diagnosis agent performs root cause analysis on detected regressions. The remediation agent recommends and, within defined bounds, implements optimizations. Results show that the agentic framework detected performance regressions with 94% accuracy and correctly diagnosed root causes in 82% of cases, compared to detection and diagnosis rates that depend heavily on manual review cadence in conventional practice. Mean time to detect performance regressions dropped from a baseline of several days under periodic manual review to under 15 minutes with continuous agentic monitoring. The agents' automated optimization recommendations, when applied, improved performance in 77% of cases without introducing regressions. The paper discusses the architecture of effective performance engineering agents, the critical role of the shared knowledge base, the appropriate boundaries of agent autonomy, and the human oversight frameworks required to deploy autonomous performance engineering safely.