Most conference talks about AI fall into one of two traps: they stay vague and aspirational, or they show a real system but never get past the architecture diagram. This is the other kind of talk.
Pipeline Detective is a multi-agent system that does automated root-cause analysis on failed CI/CD pipelines, running in production. It started as a hackathon idea and grew, step by step, into a deployed system. This talk is the build log: how the agent was actually implemented and how it evolved - the multi-agent design with Strands Agents, Amazon Bedrock and other tools, how we managed context, the ideas that worked, and the ones that didn't survive contact with production.
You'll leave with a concrete reference for how an agentic system is built end to end, and sharper instincts for which problems are good use cases for an LLM in the first place.