Standing errands it runs for you
Set it loose on a schedule
Until now you have been in the loop for every exchange. Standing errands change that: you describe a recurring job once, and the agent runs it unattended and reports back. This is levels 4–5 of the ladder — it acts on its own and tells you afterward.
- 1 Pick a genuinely recurring, low-stakes job: a morning brief, a weekly digest, a daily check.
- 2 Describe it in plain language with a schedule:
Every weekday at 8am,and send the result to me on Telegram. - 3 Let Hermes register the scheduled task (its built-in cron). Confirm it was scheduled.
- 4 Either wait for the first run or trigger a test run now.
- 5 Read the report it sends — and judge whether you would trust it to keep doing this without you watching.
A recurring task is scheduled and you have seen at least one unattended run report back to you.
When acting-without-asking is actually safe
Letting an agent act unattended sounds reckless until you see which tasks it is right for — and which it never is. Scheduled errands appear under Cron in the desktop app's status bar, and anything currently running under Agents — your at-a-glance view of what Hermes is doing unprompted.
Levels 4 and 5 — 'do it unless I say no' and 'take action, then tell me what you did' — only belong on tasks that are low-stakes and reversible: gathering, summarising, drafting, non-destructive archiving. The agent's autonomy is bounded by what its tools and skills let it touch, so the safe move is to give a standing errand only the capabilities it needs and nothing more. Post-action reporting is the safety rail: even when the agent acts alone, it leaves you a trail to read, so 'unattended' never means 'invisible'. The instant a recurring job could do something irreversible, it drops back to a level-3 gate.
- ?What makes a 'morning brief' safe to run at level 5 but 'pay this invoice' not?
- ?Why does limiting a standing errand's tools matter as much as limiting its schedule?
- ?If an unattended run did something subtly wrong, how would the report help you catch it — and how soon?
- ?Which of your recurring tasks is genuinely reversible, and which only looks reversible until it isn't?
Wire one standing errand at the right level
You will put one real recurring job on autopilot — deliberately scoped so unattended is safe.
Set up one standing errand, decide which ladder level each part of it runs at, and confirm the reporting rail works.
- 1 Choose a recurring job from your real week and define its exact trigger (when) and output (what, where).
- 2 Decide: is the whole thing level-5 safe, or does one step need a level-3 approval gate? Configure accordingly.
- 3 Give it only the tools/skills it needs — no broad access 'just in case'.
- 4 Run it (scheduled or test) and read the report.
- 5 Deliberately think of one way it could go wrong unattended, and confirm the report would surface it.
Your standing errand, the ladder level you assigned each part, its first report, and the one failure mode the report would catch.
✎The more standing errands you add, the more your assistant does while you are not looking. How will you keep track of everything it is quietly running for you — before the list grows past what you remember?