Scheduled tasks just landed in Claude Code/Desktop, and if you’ve been watching the agent space, you know why this matters: it covers a huge chunk of what people use OpenClaw for. But here’s the trap—Claude now has four different scheduling surfaces across the ecosystem, and if you pick the wrong one your task will silently die, stall on a permission prompt, or just never run. In this video I break down the three local options first: Desktop scheduled tasks (in both Co-work and the Code tab) and the CLI-only /loop command. I show how /loop works like a cron job tied to a specific Claude Code session, with non-deterministic timing (no interrupt), hard limits (3-day expiry, terminal/session dependency, up to 50 tasks), and obvious failure modes (sleep kills it). Then I contrast that with Desktop scheduled tasks: they persist to disk, survive restarts, run independently, and can use things like worktrees to avoid conflicts—yet they still won’t run if your computer sleeps, and they’re stateless unless you give them memory. My favorite pattern here is a simple state.json file that gets read/updated every run, which turns a “dumb” recurring task into a self-improving loop with lower token waste. Finally, I cover the fourth option: GitHub Actions for true always-on scheduling, with the tradeoff that the scope is basically your repo (PRs, CI failures, audits).