Most "agentic" systems fail in the same place: the agent is both the planner and the worker. Every run, it re-decides which API to call, in what order, with what arguments. That's expensive, slow, non-deterministic, and means nothing happens unless the user is steering the agent.

I built a multi-user YouTube analytics platform on the opposite premise. The agent is the configuration interface — it sets up channels, schedules, and notifications through MCP. The work itself runs as deterministic workflows triggered by cron. Configure once, run autonomously, query anytime.

This talk covers the architecture and the patterns that make it production-safe: a concrete heuristic for when a task should graduate from "agent decides" to "workflow executes"; a dual-MCP split that separates the agent's "hands" from its "brain interface"; and workflow patterns: parallel steps, typed step chaining, background execution, auto-cascade.

You'll leave knowing which parts of your agent are improvising when they should be on rails, and how to demote them without losing flexibility.