Getting started with loops

Takeaways

• Claude Code defines loops as agents repeating work cycles until a stop condition is met — not every task needs a complex loop; start simple.
• Turn-based loops (each user prompt) are the default agentic loop; encode manual verification steps in SKILL.md so the agent can self-check end-to-end.
• Goal-based loops (/goal) extend iteration with explicit, ideally deterministic success criteria and turn caps — e.g. Lighthouse score thresholds or tests passing.
• Time-based loops (/loop, /schedule) suit recurring or externally-triggered work like PR monitoring, Slack summaries, or CI fixes.
• Proactive loops compose schedule + goal + skills + auto mode for long-running streams (bug triage, dependency upgrades) without a human in real time.
• Manage token cost by matching loop type to task, capping turns, and routing judgment calls to capable models while using smaller models for routine steps.

Notes

Canonical taxonomy from the Claude Code team for designing agent loops — directly relevant to ObsidianOS skills and vault agent workflows that need verification and recurring automation.

Open questions

• Which vault skills (e.g. /vault-health, /kb-triage) benefit from /goal vs turn-based loops with encoded verification?