How to design AI agent loops: schedules, goals, and subagents in Claude Code and Codex

What It Covers

This episode demystifies AI agent loops — scheduled, goal-based, and hook-triggered automations in Claude Code and Codex — explaining how to design agents that prompt themselves autonomously, deploy sub-agents for parallel work, and validate outcomes without human input, using practical product and engineering workflow examples.

Key Questions Answered

• •Loop Types — Three Distinct Formats: Agent loops take three forms: heartbeats (recurring intervals like every 5 minutes), crons (fixed schedule like every Friday at 9AM), and hooks (event-triggered via webhooks or internal lifecycle events). A fourth type — goal loops — runs continuously until a defined success condition is met or the agent becomes blocked, then stops automatically.
• •Goal Loop Precision: Goal-based loops require explicitly defined evaluation and success criteria in the prompt. Vague goals burn tokens without useful output. OpenAI publishes a dedicated guide for writing Codex goals. A reliable pattern: prompt the sub-agent with a specific, measurable validation target against a defined branch or dataset before the loop begins executing.
• •Sub-Agent Architecture: Both Claude Code and Codex support spawning sub-agents from a parent loop. The parent identifies tasks, then delegates each to a dedicated thread with its own goal loop for validation. In a live demo, a Friday automation scanned recent PRs, generated missing skills, and spawned named sub-agents — Gauss, Galileo — each pursuing independent validation goals concurrently.
• •Loop Cost Management: Loops with loose validation criteria or wide-ranging scope burn tokens rapidly. Goal loops are especially expensive because agents iterate until thresholds are self-assessed as met. Monitoring both cost and execution efficiency is necessary from day one. Applying loops only to well-scoped, repeatable tasks with precise success criteria reduces unnecessary token consumption significantly.
• •Practical Loop Design Framework: Designing a loop mirrors writing a job description for an employee. Define the schedule or trigger, the specific task, the tools available (GitHub, Slack, Google Calendar connectors), and the done condition. A daily PR aging review loop in Claude Code checks for PRs open over 12 hours, babysits merge checks, and posts Slack alerts — all without manual prompting.