An agent demo takes a weekend. An agent integrated into a vertical SaaS product serving 50,000 users takes a serious harness. At Productive we rebuilt ours three times. This talk is a deep dive into the architecture and patterns that survived. Why moving all reasoning into the model drastically improves performance. How the agent explores a huge SaaS domain on demand instead of carrying it in the prompt. How to properly write instructions & prompts and the most common mistakes people make there. How the agent manages its own context as a scarce resource. And the production loop: evals, LLM-based session triage, an agent that vents its frustrations to our Slack and other observability tricks. Real examples and code with a fun interactive demo at the end.

Recorded as part of the Foam Stage day-1 live stream; the embed starts at the beginning of the talk.