Sherlock: Building an Agent That Becomes an Expert on Any Domain

What does it actually take to build an AI agent that can work unsupervised, handed a task and trusting it to come back with something useful? This talk cuts through the hype with an honest, battle-tested account of building Sherlock, a multi-agent system designed to reason over complex domains, extract knowledge from diverse sources, and autonomously develop the tools it needs to do so.

The presentation will cover building a production-grade AI system, including the lessons learned and challenges that shaped its architecture: how memory works (and breaks) at scale, when to trust a model's judgment vs. enforce hard rules, why deterministic and probabilistic execution need to be treated as fundamentally different things, and how multi-agent coordination introduces new failure modes even as it solves old ones.

You'll leave with a clearer mental model of how production-grade agent systems are actually built, what keeps them from working, and what hard-won lessons look like in practice.