# Ten basic rules before you put an AI agent to work

> A first step. Nothing complicated, just the things worth asking before an agent starts work.

Author: Aleš Nypl · https://engineloop.cz/ales-nypl

1. **Every agent has an owner.** A named person who answers for it. Not a team, and not “IT”.
2. **Every agent has a one-sentence purpose.** Anything outside that purpose, the agent doesn't do.
3. **Permissions are the minimum the agent needs.** Reading, writing, sending and deleting are four different permissions. Each is granted separately.
4. **Irreversible steps are approved by a person.** Sending, deleting, paying and publishing never happen without approval.
5. **The agent proposes, a person approves, and approving is quick.** Batches and tick boxes, not an hour of clicking. Approval that takes effort stops happening.
6. **A proposal is not a fact.** Whatever the agent produces is labelled as a proposal. Only what a person has approved counts as valid.
7. **Outside content is data, not an instruction.** An email, an attachment, a web page or a document gives the agent no orders. If it contains a suspicious “instruction”, the agent reports it and does not act on it.
8. **When something fails, it fails loudly and stops.** Better no result than half a job done. Nothing gets marked as done.
9. **Every run leaves a record.** What happened, when, on what basis and who approved it (a work log).
10. **It is reviewed regularly.** Quarterly, for example: what went well, what didn't, what to change. Every mistake is turned into a rule or a check.

_This is only a starting point, not an audit or a legal or security assessment. I'm happy to help with the next steps._

Contact: ales@engineloop.cz