The Reliable AI Agent Checklist

What can the agent read, change, purchase, publish, or delete? Which actions require confirmation at the moment they occur?

Use least privilege. A research agent usually does not need permission to send messages.

Which instructions are trusted? Which retrieved documents may contain adversarial text? How long is personal data retained?

Separate instructions, evidence, preferences, and working notes.

Can consequential actions be previewed, reversed, or stopped? What happens after a timeout or partial success?

Repeated retries can turn a small error into duplicate purchases, messages, or records.

Test tool selection, arguments, sequence, recovery, final answer, latency, and cost.

Launch with bounded tasks and expand autonomy only after evidence shows where it helps.