All guides
Free blueprint
For product, engineering and IT leaders at U.S. software companies letting an AI agent change records, send messages or move money in their own software
Your agent retried a payment. The customer got paid twice.
Before your AI agent touches money: limits, approvals, undo
A payment timed out, the agent retried, and the customer got paid twice. Decide what an agent may do alone, what waits for a person, and what happens when a step fails.
This answers
- Which actions may an agent take alone, and who approves the rest?
- What happens when a payment times out or a step fails halfway?
- How do we stop text the agent reads from steering what it does?
Blueprint · 8 pages, PDF · Proposed approach
What is inside
- 1Harbor Billing: one refund, one approval, one safe retryA filled example
- 2Three yes answers make a run dangerousStep 1 · Mark each agent run
- 3Five controls a fooled model cannot switch offStep 2 · Controls in code
- 4Decide what happens before it goes wrongStep 3 · Limits and failures
- 5Agents that prepare the action. A person confirms itFrom our work
- 6Bring this page to the Product ReviewYour notes for the call
You also get
- A sample refund with one approval and one safe retry
- Three questions that mark a dangerous agent run
- Five controls a fooled model cannot switch off
- What to do when a limit is hit, a call times out or a step fails halfway
Preview
The first 3 of 8 pages

Page 1 of 8

Page 2 of 8

Page 3 of 8
- 5 more pagesFive controls in code, limits and failures, proof from our work and a page to bring to the call.Get the full guide
