AI-Wiki

How to write AI agent loops in Claude Code and Codex

7 min read

How to write AI agent loops in Claude Code and Codex

I break down every loop type from scratch—what a heartbeat, cron, hook, and goal loop actually are, when each one fits, and the five things any effective loop needs before it touches production. Then I build two live loops: a daily aging-PR reviewer in Claude Code that schedules itself at 10:15 a.m. and spins off its own subagents, and a weekly skills-identification loop in Codex that spawns goal-based subagents to validate its own output in real time.

— Channel description, How I AI (episode hosted by Claire Vo)

A ~29-minute explainer episode of the How I AI podcast, hosted by Claire Vo (creator of ChatPRD), demystifying the practitioner term “loop” — an automated prompt, where an agent prompts itself rather than waiting for a human to type each turn. The episode’s framing claim, borrowed from a tweet by “Pete at OpenClaw,” is deliberately provocative — “prompts are out and loops are in” — but Vo’s substantive contribution is the opposite of hype: a plain-English taxonomy of loop types, a checklist of what every effective loop needs, a management mental model for designing loops, two live builds, and an honest set of warnings about cost. It sits at the product-leader / accessibility altitude of the wiki’s agentic-engineering cluster, complementing the engineer-altitude treatments by Lopopolo and Osmani.

TL;DR

  • A loop is just an automated prompt. The reframe: you do not need human fingers typing a prompt for an agent to do work. Automated prompting predates AI — we already triggered automations on schedules, crons, and webhooks; loops apply the same triggers to an LLM agent.
  • Four ways to automate a prompt — (1) heartbeat (recurring interval: “every 5 min, check for a new Jira ticket”); (2) cron (fixed schedule: “at 9:00 a.m.”); (3) hooks (internal-lifecycle events like tool-called / session-started, or external webhooks like “every email received”); (4) goal — the new first-class loop shipped in both Claude Code and Codex: set an outcome, run the agent until the outcome is measured-and-validated or the agent is blocked.
  • Five things every effective loop needs (crediting Addy Osmani’s loop engineering article): work trees (isolate each agent’s work in a sandbox), skills (reusable ways to do common tasks), plugins/connectors (the tools the agent can reach — GitHub, Slack, Google Docs/Calendar), subagents (federate work off the main thread, especially validation), and state tracking (a to-do list — markdown or Linear).
  • The “onboarding an employee” framework. Designing a loop = designing a job. Imagine onboarding an EA, a support agent, or a software engineer and writing their recurring duties: “Every Friday, review my calendar and Slack me the follow-ups” is a loop; “every PR, iterate until all checks are green and it’s ready to deploy” is a goal loop.
  • Two live builds — (1) in Claude Code, a daily aging-PR review routine (10:15 a.m., local) that finds PRs open >12h, spins up a subagent thread to babysit each until merge checks are green, and Slacks the team about the rest; (2) in Codex, a weekly skills-identification automation that reads recent PRs/commits, proposes missing skills, and spawns goal-based subagents to validate each new skill against the base branch — a loop that designs and runs its own sub-loops.
  • Warnings. Loops burn tokens — goal loops especially, because they run validation until a threshold the agent decides is met; a thin validation criterion is a money pit. And goal/loop prompts are harder to write well than conversational prompts — success and evaluation criteria must be precise (Vo points to OpenAI’s Using Goals in Codex guide, and often has Codex write its own goals).

Why this matters to the wiki

This source is the wiki’s first product-leader-altitude, non-engineer-facing treatment of the self-prompting loop, and it does three useful things for the agent-harness / agentic-engineering cluster:

  1. A clean loop taxonomy. Heartbeat / cron / hook / goal gives the wiki a vocabulary for how an agent gets prompted without a human turn. The first three are pre-AI automation triggers re-pointed at an LLM; the fourth — the goal loop — is the genuinely new primitive, now first-class in both Claude Code and Codex. This sharpens the agent-harness page’s account of the reason–act–observe loop by naming the trigger layer that kicks the loop off.
  2. The goal loop as a named primitive. Vo’s definition — run against a measurable outcome until validated or blocked — is the product-facing name for the long-running agent that Anthropic engineers vendor-side and that Osmani calls a Ralph loop. Three independent vocabularies converging on one construct is exactly the kind of cross-source ratification the wiki tracks.
  3. The “onboarding an employee” management frame. Designing loops as jobs to be done is a memorable bridge between the automation-vs-augmentation discussion and the day-to-day practice of agentic-engineering — and it operationalises the recurring “manager of agents” motif (cf. Jassy on the changing role of managers).

The four loop types

TypeTriggerExample from the episode
HeartbeatRecurring interval”Every 5 minutes, check for a new Jira ticket; if one exists, start a coding agent to triage and fix it.”
CronFixed/defined schedule”At 9:00 a.m.” — more precisely scheduled than a heartbeat’s regular cadence.
HookInternal lifecycle event (tool-called, session-started/reset) or external webhook”Every time I receive an email, fire a webhook and kick off an agent.”
GoalRun until outcome validated or blocked”Iterate on this PR until all lints/checks are green and it’s ready to deploy.” — the new first-class loop in Claude Code (/goal) and Codex.

Vo’s synthesis: a loop kicks off an automatic prompt and keeps it going until the job is done — done either because time’s up (scheduled loops) or because the job’s done (goal loops) — then empowers the kicked-off agent with the isolation (work trees) and tools (connectors, subagents) it needs.

The two live builds

  • Build #1 — Daily aging-PR review (Claude Code routine). Runs locally, daily at 10:15 a.m., scoped to one branch. Instruction: find PRs open >12h on the ChatPRD app; for any that can be babysat, spin up a thread to watch the PR until merge checks are green; otherwise Slack the product channel (“be mean”). Demonstrates a scheduled loop that spawns subagent threads — not all work happens in the master thread.
  • Build #2 — Weekly skills-identification (Codex automation). Built from a Codex template, runs Fridays weekly. Reads recent PRs/commits, proposes missing skills (especially around internal CLI/MCP validation tools), then spawns a goal-based subagent per skill to validate it against the base branch. Vo watches Codex assign named subagents (it scaffolds its own memory) and confirm a candidate — “a loop with subagents that generates its own loop.” This is the episode’s “mega loop”: a scheduled loop → that sets up a team → that does work on goal sub-loops.

What was actually ingested

Full ~29-minute episode transcript (auto-generated English captions, ASR-cleaned for proper nouns: OpenClaw, ChatPRD, Codex, Claude Cowork, chat-smoke CLI). Two sponsor reads (WorkOS, Runway) are present in the transcript and were not treated as content. Host identity (Claire Vo), tool names, and the Osmani/OpenClaw references were confirmed against the video description’s link block, not inferred.

Linked entities and concepts

(Note: Claire Vo and How I AI were promoted to entity pages on this ingest — both reached their second source here, the first being the Marily Nika episode.)

Relationships

Supports

  • Ryan Lopopolo — Harness Engineering: How to Build Software When Humans Steer and Agents Execute (AI Native DevCon) — Two June-2026 practitioner agentic-engineering talks at different altitudes. Lopopolo is engineer-altitude — designing harnesses so agents execute at production scale (~1M LOC, 0 hand-written lines). Vo is product-leader-altitude — making the *self-prompting loop* legible to non-engineers via the four loop types and the onboarding-an-employee framework. Both land the same load-bearing move: the human stops typing each prompt and instead designs the automation that prompts the agent.
  • Agent Harness Engineering — Vo explicitly cites Osmani's *loop engineering* article as the best breakdown of what an effective loop needs, and reproduces its five-element checklist (work trees, skills, plugins/connectors, subagents, state). Osmani's *Ralph loop* and Vo's *goal loop* are the same self-prompting-until-done primitive named from the engineer side vs. the product side.
  • Effective harnesses for long-running agents — Vo's *goal loop* — set a measurable outcome, run the agent against it until validated or blocked — is the product-facing name for the long-running agent Anthropic engineers vendor-side. Vo's warnings (goal loops burn tokens until a too-thin validation criterion is satisfied; precise success criteria are everything) are the practitioner echo of Anthropic's reliability concerns for long-running loops.

Graph View

Backlinks