# OpenClaw Standing Orders Standing orders are [[OpenClaw]]'s mechanism for **persistent operating authority**: a written mandate that gives an agent permission to do *X* under conditions *Y* without re-asking each time. Defined primarily in `AGENTS.md` files inside a workspace, they make autonomous loops legitimate and auditable rather than ad-hoc. The pattern matters because chat-driven agents are otherwise stuck in synchronous request/response. Standing orders unlock the "while I sleep, do this" mode that makes overnight coding supervision, automated triage, and scheduled briefings safe to deploy. ## What a Standing Order Specifies - **Scope** ; what the agent is authorized to act on (which repos, which channels, which tools) - **Triggers** ; what fires the order (cron schedule, inbound event, heartbeat, webhook, threshold) - **Approval gates** ; for which actions does it still need a human in the loop - **Escalation rules** ; what to do when uncertain or blocked - **Reporting cadence** ; when and where to surface results ## Execute-Verify-Report Pattern Programs running under standing orders follow the documented **Execute-Verify-Report** loop: 1. **Execute** ; do the work as instructed 2. **Verify** ; check that the work meets the stated success criteria 3. **Report** ; surface what was done, what wasn't, and why ; in the agreed channel This is the discipline that prevents silent drift. A standing order without a Report step is a black hole. ## Time-Based Triggers Time-based triggers are enforced via **cron jobs**: a separate scheduling layer that wakes the agent at `*/15 * * * *`, `0 7 * * 1-5`, etc., and dispatches the matching standing order. Standing orders define *what*; cron defines *when*. Other dispatch primitives include heartbeat, polls, and webhooks. ## Risk Posture Standing orders are powerful, which means the boundary work matters. Recommended posture: - Start narrow ; small scope, tight schedule, explicit approval gates - Add a **command-logger** [[OpenClaw Hooks|hook]] for audit - Run risky orders in **sandboxed sessions** when possible - Review the report channel before widening scope ## Composition A typical autonomous routine layers four primitives: 1. A **standing order** with the policy and triggers 2. A **cron job** firing at the right cadence 3. **Sub-agents** doing the parallel work in the background 4. **Hooks** observing each step for audit ## References - Standing Orders: https://docs.openclaw.ai/automation/standing-orders ## Related - [[OpenClaw]] - [[OpenClaw Hooks]] - [[OpenClaw Sub-Agents]] - [[OpenClaw Skills]] - [[Claude Code Routines]]