adsagents Agent Scheduled Tasks
- Classify the job as reminder/heartbeat, auditable execution, or consequential execution.
- Resolve timezone, cadence, destination, stable scope, allowed tools, evidence, retry budget, and stop conditions before creating it.
- Use the actual scheduler interface once, then read back its stored contract.
- Run now only when supported and record the scheduler plus downstream evidence separately.
- Never put credentials in the prompt, guess permissions, or treat schedule creation as execution proof.
- Treat a matching delivery prepare as the live preflight. Consume inline verification first and use the advertised native operation fallback only when recovery remains pending; a missing optional selector before mutation does not disable read-only schedule evaluation.
- For explicit Meta candidates, batch up to 50 exact IDs with
overview_get_live_configs; correlate only byentity_typeandentity_id.
Read scheduled-task-contract.md only when designing, creating, updating, debugging, pausing, deleting, or proving a scheduled task.