A Retool engineer explains how the company built production AI agents by leveraging their existing Temporal-based workflow engine. The talk covers the core challenges of production agents (durable execution, context management, unreliable tools, human approval, observability) and how Temporal primitives already solve them. It details three iterations of Retool's workflow execution engine — from naive topological execution to a parallelized static DAG pre-planner — achieving 5x fewer activities, 8x faster execution, and $9M/year in savings at 10M+ daily workflow runs. The agent architecture is revealed to be a dynamic state machine built on the same Temporal workflows and activities, with human-in-the-loop implemented in ~50 lines using Temporal signals. Practical patterns for building your own agentic infrastructure are shared, including child workflows for sub-agents, event history for observability, and schedules for recurring agents.