Trust, but Verify: Building a Document Agent That Survives Production

Let’s ship an agent into production. You’ll start with the easy part: a document comes in, an agent reads it, pulls the fields, hands off the result — the kind of thing you could wire together yourself from an agent framework and a workflow engine. We’ll build that starting point live, then do the part the demos always skip: keeping it alive in production.

Real documents don’t match your template, and the agent that worked at design time answers differently each time. Over two hours you’ll grow a simple document flow into a governed agentic workflow: adding an agent with inspectable traces, writing evals that validate its output and trajectory, and catching a regression before it ships. Then you’ll publish it to run on durable, recoverable orchestration. You’ll leave with a running agent and a reusable eval set you can point at your own work.

What you’ll walk away with - A document-processing agent you built and watched run - Output and trajectory evals you can reuse on your own projects - A feel for what “production-grade” actually takes: durable execution, governance, recovery