agent-loop

Agent Loop

Use this skill for multi-step coding tasks that need a bounded agent loop instead of a single local step.

When to use

• The user says loop, agent loop, autonomous, keep iterating, or mentions Claude Code.
• The work will likely require repeated inspect/edit/test cycles.
• A turn cap or explicit approval boundary is useful.

Defaults

• --max-turns 8
• --approval-mode on-write
• --model gpt-5.4
• --reasoning-effort high

Important:

• Agent Loop is bounded. It should stop when the task looks complete, when it hits its turn/time budget, when it pauses for approval, or when an error occurs.
• In fallback mode without OPENAI_API_KEY, Agent Loop runs one codex exec session. That session may finish on the first turn for short read-only tasks.
• A stop reason like completed_via_fallback means the task completed through the fallback backend; it does not mean the run failed.

Shorthand budgets are also supported:

• 10m = 10 minutes
• 1h = 1 hour
• 5t = 5 turns

Invocation

Prefer invoking the bundled skill directly:

$agent-loop <task>

Or with a shorthand budget:

$agent-loop 10m <task>

Some Codex builds may also surface the plugin command definition. If yours does, the equivalent command is:

/agent-loop:run <task>

Onboarding helpers

• --doctor checks that the plugin is installed and ready in Codex.
• --demo runs a safe read-only first-run example and suggests next commands.
• 10m, 1h, and 5t work as shorthand budgets before the task text.

Approval modes in plain English:

• on-write = safe
• never = hands-off
• always = review-everything

Approval behavior

• on-write auto-runs read-only shell inspection commands.
• on-write pauses before write-like shell commands and all patches.
• Use --resume <state-file> --approve-pending only after the user explicitly approves the paused action.

Run artifacts

The runner stores logs and resumable state under: