A loop is an automated prompt. That framing, introduced in this episode, strips the concept down to its functional core and then builds it back up into four distinct types: heartbeat, cron, hook, and goal. Each has a specific fit. Using the wrong one wastes tokens before your agent does anything useful.
Two live builds anchor the theory. The first is a daily PR-review agent in Claude Code, scheduled at 10:15 a.m., that identifies aging pull requests and spins off subagents to alert teams. The second is a weekly skills-identification loop in Codex that spawns goal-based subagents to validate its own output in real time. Both demos expose the five prerequisites every loop needs before production: work trees, skills, plugins and connectors, subagents, and state tracking.
The section worth the full watch starts at 25:28, where two specific warning signals predict whether your loop will get expensive before it gets useful. Goal-based loops are flagged as the hardest type to write correctly, and the episode names exactly where most implementations burn budget without producing output. The 'onboarding an employee' mental model at 9:26 reframes loop design in a way that makes the five prerequisites feel structural rather than optional.
[READ ORIGINAL →]