SE Radio 719: Birol Yildiz on Building an Agentic AI SRE

In this episode of Software Engineering Radio, host Kanchan interviews Birol Yildiz, CEO and co-founder of Islet, a SaaS company building an AI-powered SRE (Incident Response) system called AI SRE. The conversation dives deep into the architecture, evolution, and philosophy behind creating an agentic AI system that autonomously performs root cause analysis for production incidents. Yildiz explains how the team evolved from a rigid, prescriptive approach using complex workflows and vector databases to a minimalist, agentic model that relies on reasoning loops and 'agentic search'—using simple command-line tools like grep, jq, and bash to query data without overloading context. The AI SRE is designed to complete root cause analysis in under four minutes, a dramatic improvement over manual processes that can take 10 to 60 minutes. The discussion covers key components: orchestration, knowledge layer (using plain-text long-term memory instead of vector databases), evaluation via semantic tests and LLM judges, and the use of sub-agents and forks to manage context. A real-world example illustrates how the AI SRE diagnosed a self-inflicted incident caused by an overly broad network policy during a penetration test, demonstrating its ability to handle ambiguous, novel problems beyond standard runbooks. The episode also explores guardrails, autonomy, GDPR compliance, and the future of AI agents, with Yildiz cautioning against over-engineering and advocating for simplicity, full context control, and letting the model decide the 'how' while humans define the 'what'.