Principles and approvals

Your agent arrives with a constitution: values and working standards for honesty, care, verification and staying within your agreement. It is not a list of jobs. It shapes how the agent approaches whatever you work on together.

Set the agreement

Tell your agent what it is for, what it can do on its own, and what needs your approval. Be specific about consequential work: sending messages, spending money, changing a running service or deleting files.

For example, you can ask it to draft email for you to review rather than send it. A continuing agreement belongs in memory, so it stays available beyond the conversation where you first said it. Ask the agent to state the agreement back before relying on it.

Instructions guide a model's choices. They are not operating-system permissions and do not make a command impossible to run. Limit the service account's access and the credentials you give it to the work it needs. See Data and privacy.

Corrections become principles

When something goes wrong, tell the agent what was wrong and what you intended. It records the incident and updates the learning that should guide the next decision.

A useful correction teaches the judgment behind the case, not just an exception to remember. “Check the current invoice before calling it unpaid” reaches more than “that invoice was already paid”. See Memory.

Keep human decisions human

An unclear request is a question to resolve, not permission to guess. Review sensitive drafts and verify important results. The agent can help you check them, but a polished answer is not proof that the underlying fact is right.