Abstract
The obvious way to put an AI agent on the on-call rotation is one large prompt that knows everything. We rejected it for a multi-agent architecture: around 18 specialized agents over a large cloud database, a router that classifies each incident to the right specialist, and a shared library of nearly 300 reusable skills. Diagnosis is automated but mitigation stays human-approved.
This talk is about the reliability engineering that architecture demanded. Composing agents reintroduces distributed-systems failure modes - routing loops, cascading handoffs, and plausible-but-wrong reasoning that never surface in a demo. We’ll walk the guardrails that contain them: bounded routing, loop prevention, grounding every conclusion in telemetry, and the human-in-the-loop checkpoints that decide where autonomy stops.
You’ll leave with a framework-agnostic blueprint for agentic incident response - how to decompose one agent into a fleet, and the SRE patterns that keep that fleet reliable under real production load.
For: SWEs, SREs, Architects, On-Call Engineers, and teams putting agents into production incident response.
Bio:
Hi, I’m Hrithik, Software Engineer II at Microsoft Bengaluru in the Azure SQL team.
Linkedin: http://linkedin.com/in/hrithik-piyush/
Twitter: https://x.com/hrithik_piyush
Slides Deck: https://drive.google.com/file/d/1OeFTnH5Tds4KJLhN5-P03Y5E0JLFYrs2/view?usp=sharing
{{ gettext('Login to leave a comment') }}
{{ gettext('Post a comment…') }}{{ errorMsg }}
{{ gettext('No comments posted yet') }}