Detailed description

AI agents for SRE are moving from experiments to production, but reliability, observability, and safety are still major challenges. In this session, we present our autonomous SRE diagnosis agent built with Microsoft Agent Framework, Magentic orchestration, Memgraph, k8s, OpenSearch, and OpenTelemetry.

We explain why we chose graph-based orchestration for long-running diagnostic workflows and share practical lessons from building the system: controlling context explosion, parallelizing investigative agents, enforcing structured outputs, and deduplicating noisy alerts. We also show how graph databases enables operational memory by correlating incidents and reusing past diagnostic insights.

Finally, we discuss safety boundaries for autonomous operations and the challenge of evaluating non-deterministic AI-driven SRE systems using ITBench and internal experiments.

Takeaways

• Tradeoffs between Pregel, A2A, and event-driven agent architectures.
• Why orchestration matters more than raw model reasoning.
• Practical techniques for avoiding context explosion in observability systems.
• Using Memgraph to correlate similar incidents and reuse operational knowledge.
• Magentic-One as an alternative to Group Chat and ReACT
• Safety and auditability patterns for autonomous SRE agents.

Target audience

• SREs and DevOps Engineers
• AI Engineers
• Architects designing autonomous operational tooling
• Engineering leaders evaluating AI for production infrastructure