How Prisma Built An AI SRE
Agents get harder to operate the moment they touch real production data. The advice usually stops at "add a sub-agent" and "pick a good model," and neither survives contact with an on-call rotation. In this workshop, Tyler Benfield from Prisma will walk through Sherlog, the Slack-native operations health agent Prisma runs on its own telemetry. He’ll explain how the team uses sub-agents in production and why it chose open-weight models for the entire system. We’ll cover: How Sherlog’s Axiom query executor uses a sub-agent to process high-volume log data without filling the main agent’s context, and how it summarizes the results Why speed, cost, and separation of concerns are often poor reasons to introduce a sub-agent, with examples of tasks Prisma kept in the main agent How Prisma assigns models by role, giving reasoning, code review, data analysis, and classification different capabilities through a shared configuration layer How Prisma evaluated cloud and open-weight models, including the tradeoffs behind its move to a fully open-weight system Expect a live demo, real code, and plenty of time for questions. Hosted by: Brandon Barros, Founding Product Advocate at Mastra Tyler Benfield, Platform Architect at Prisma