Running alone
Unattended means the agent runs whether or not anyone is looking. It is the point of an always-on agent and the moment it becomes a system with real consequences, because everything it does at 3am it does without a person to catch it. Here is what has to be true before that is a good idea.
Direct answer
Four things: a host that stays online without a person, credentials scoped to only what the agent needs, a hard spend ceiling that stops rather than warns, and actions that are either reversible or gated behind an approval. Missing any one of them turns an ordinary failure into an incident nobody sees until later.
| Action | Unattended? | Why |
|---|---|---|
| Read and summarise sources | Yes | Reversible, no external effect |
| Draft a reply for review | Yes | Nothing leaves until you send it |
| Flag and prioritise an inbox | Yes | Changes labels, not outcomes |
| Write a report to a workspace | Yes | Contained and reversible |
| Send external email | Gate it | Cannot be unsent |
| Make a purchase | Gate it | Real money, hard to reverse |
| Delete anything | Gate it | Often unrecoverable |
| Post publicly | Gate it | Reputational, and permanent |
An unattended agent on a laptop is a scheduled agent with a single point of failure wearing shoes. The lid closes, the machine sleeps, the schedule passes and does not catch up. A VPS or a managed environment removes that particular dependency; only one of the two also removes the server.
The temptation with an unattended agent is one broad key, because it makes everything work first time. It also means any mistake the agent makes is as wide as that key. Scope is the cheapest safety measure available and it costs one afternoon.
An attended agent that starts looping gets closed by whoever is watching. An unattended one does not, and the failure compounds for as long as it goes unnoticed. A warning at 80 percent is useful; a hard stop is what actually bounds the damage.
This is the judgment call, and it is worth making deliberately rather than discovering. Sort what the agent can do into things you can undo and things you cannot, and treat the two differently.
The distinction matters and gets lost in the marketing. Unattended is about execution: the agent runs without a person present. Supervised is about accountability: a person still reads what it produced and remains responsible for it. You want the first without giving up the second.
Agents run in managed, isolated environments that stay online without a person, credentials live in an encrypted vault and are injected at runtime, the credit balance carries a hard stop rather than a warning, and the Hermes runtime supports an approval mode where an action waits for human sign-off before it happens.
The honest answer
Yes, but less than you think and differently than you expect. Which parts of running an agent can be automated away, and which stay a human responsibility.
Read guideMoney
How AI agents run up unexpected bills: retry loops, growing context, oversized tool payloads. Why alerts arrive too late, and what a real hard stop looks like.
Read guideThe homework, itemised
Running an AI agent is easy. Keeping it running is the homework: restarts, key rotation, silent failures, runaway spend. The whole list, and who does it.
Read guideUnattended describes when it runs: on a schedule or trigger, without a person present. Autonomous describes how much it decides for itself. An agent can be unattended and tightly scripted, or attended and highly autonomous. Conflating the two is how people end up granting broad permissions they did not intend.
Deploy a managed agent environment, configure the runtime, and keep the agent online without Docker, VPS setup, or server maintenance.
Get started