Agent Interrupt Rules

Decide the boundaries before handing over a job.

Adapted from a three-bucket framework shared by Allie K. Miller: sort agent work by how bad it is if the agent gets it wrong.

Sort the jobs

Bucket 1 · Easy to undo, limited impact: handle within explicitly approved limits. Examples: label newsletters, prepare a private briefing. Your job: ___ Limits: ___

Bucket 2 · Review before acting in your name: draft and wait for approval. Examples: client reply, meeting invitation, team message. Your job: ___ Approver: ___

Bucket 3 · Hard to undo or consequential: bring it to the owner before acting. Examples: payments, contracts, HR or legal matters, or actions involving anyone you know personally. Your job: ___ Owner: ___

Classify the actual action and permissions, not just its title. Filing a receipt can still expose private information if the destination is wrong.

Paste-in instruction block

For this job, use only these approved inputs and permissions: [list].

Bucket 1 actions allowed without asking: [specific list and limits].

Bucket 2 actions: [list]. Draft only and wait for my explicit approval before sending or changing anything.

Bucket 3 actions: [list]. Stop and bring the decision to [owner].

If an action is unlisted or ambiguous, ask before proceeding.

Tell me when you can’t finish a job: what failed, what is complete, what remains, and what you need. Use [configured alert route]. Do not silently treat partial work as complete.

Worked example · fictional

A customer asks for a refund. Summarizing the request privately is allowed; drafting the reply needs review; issuing the refund is reserved for the owner. No reply or payment happens while approval is pending.

Monthly review

Review date: ___ Job: ___ Errors and exceptions: ___ Proposed boundary change: ___ Owner approval: ___

Move a job toward more autonomy only after reviewing evidence and explicitly approving new limits. Instructions alone do not enforce permissions or install alerts; configure and test those separately.