Podcast
How to Design AI Agent Loops- Schedules, Goals, and Subagents in Claude Code and Codex
How I AI
- Three Ways To Automate Agent Prompts
- Claire distinguishes manual message prompting (human-initiated chat turns) from automated loops that prompt agents without human input.
- She defines three loop form factors: heartbeats (regular intervals, e.g., every 5 minutes), crons (time-specific schedules like 9am or Sundays), and hooks (triggered by internal events like a tool call or session start).
- Heartbeats are useful for frequent polling tasks (e.g., check JIRA every 5 minutes and triage new tickets).
- Crons are for less frequent, fixed-timing tasks (daily reports, weekly jobs).
- Hooks tie prompts to lifecycle events, enabling reactive, event-driven agent runs. (Time 0:03:03)
- Loops Are Automated Prompts Not A New Paradigm
- A loop is simply an automated prompt that runs without human typing, triggered on schedules, hooks, heartbeats, or goals.
- Claire Vo lists heartbeat (regular), cron (fixed time), hooks (events), and goal loops (run until validated) as the core forms. (Time 0:03:05)
- Build Loops With Five Foundational Components
- Build loops with five foundations: work trees, skills, plugins/connectors, subagents, and state tracking.
- These keep agent work isolated, reusable, connected to tools, delegable to subagents, and trackable as a to-do list. (Time 0:06:24)
- Design Loops Using An Employee Onboarding Mindset
- Design a loop like you’re onboarding an employee: define the job, schedule, and desired outcomes.
- Examples: weekly EA calendar review, hourly issue triage, or goal-based PR review until lints pass. (Time 0:09:30)
- Daily Aging PR Reviewer That Babysits PRs
- Claire built a Claude Code routine named Daily Aging PR Review to find PRs older than 12 hours and either babysit merge checks or post a Slack alert.
- It runs daily at 10:15am, uses GitHub and Slack connectors, and can spawn subthreads to watch PRs. (Time 0:13:26)
- Weekly Skills Identification That Spawns Validating Subagents
- In Codex Claire used a template to scan recent PRs for missing skills, then spawned subagents that run goal loops to validate each identified skill.
- The automation created chat threads that ran CLI smoke tests against the base branch to prove the skill works. (Time 0:20:17)
- Avoid Token Burn With Precise Goal Validation
- Monitor loop costs and write precise goal prompts to avoid burning tokens.
- Goal loops especially must have strict, measurable validation criteria or they will run long and expensive validations. (Time 0:25:36)