Proposed
The agent has described or prepared work. No external action is implied.
CONTROL / EXECUTION / PROOF
Keep consequential actions behind a clear yes. A plan can be useful without being permission to change the outside world.
Approval should identify the exact action and scope. This guide describes the boundary; enforcement still depends on the specific tool and integration.
01 / KEEP THE STATES SEPARATE
When a status skips a step, people can mistake intent for action or action for success. Name what happened—and what is still unknown.
The agent has described or prepared work. No external action is implied.
A person authorized one clearly described action, target, and scope.
A tool call was attempted. Preserve the returned status; it may be pending or failed.
A separate source or read-back observed the intended outside result.
02 / DECIDE WHAT NEEDS A YES
Teams often require a fresh approval before an action that affects money, access, production systems, or another person. These are policy examples, not a claim that every tool uses the same gate.
Messages, posts, offers, customer notices, and public changes that leave the workspace.
Purchases, payments, trades, transfers, and other actions with financial consequences.
Production settings, customer data, repositories, records, or other hard-to-reverse state.
Permissions, credentials, scopes, or access that lets another person or process act.
03 / MAKE CONSENT SPECIFIC
Show the reviewer the actual proposed action—not just a goal or a vague “continue?” prompt—before they approve it.
State what will happen, including the tool or service that will do it.
Name the recipient, account, record, environment, or resource affected.
Show amounts, content, timing, quantity, and any important boundary.
Make clear who is approving and whether they are allowed to authorize it.
Explain how long approval lasts and what else the action may change.
Say what can be stopped, reversed, or escalated if the result is wrong.
04 / AFTER APPROVAL
Approval is a boundary, not a shortcut around verification. Keep the handoff narrow and return an honest status.
| Step | Good practice | Do not claim |
|---|---|---|
| Re-check | Compare the approved action with the request about to run. Stop if the scope changed. | That an earlier, broader approval covers every later variation. |
| Execute | Record which tool ran, when it ran, and whether the call returned success, failure, or pending. | That a drafted plan or queued job has already changed outside state. |
| Verify | Use a provider response, read-back, settlement reference, or target-system check when available. | That a local “success” response proves delivery or final settlement by itself. |
| Report | Show the evidence, remaining uncertainty, and the next safe action. | Completion when the result is blocked, ambiguous, or unconfirmed. |
05 / RECEIPTS AND EVIDENCE
Keep the record tied to what the system observed. When the result matters outside that system, add an independent confirmation.
Verify the outside state separately. A verifiable record can support an audit of what the system recorded. It does not turn missing upstream evidence into proof of delivery, settlement, or publication.
06 / COMMON QUESTIONS
No. A plan describes proposed work. Execution requires a tool or provider call; confirm the resulting state separately.
No. Approval authorizes a specific next step. It does not show that a request ran or that an external system applied it.
Treat a changed recipient, target, amount, scope, or side effect as a new request and obtain approval for the revised action.
A receipt can document the call and the result observed by the system that issued it. It does not independently prove an unobserved delivery, payment, publication, or state change.
Often, yes, if the data access is authorized and no consequential side effect occurs. Access and privacy rules still apply; confirm the tool's actual scope.
No blanket guarantee is implied. Approval behavior depends on the specific integration and action. Check the current tool contract and its confirmation mechanism before relying on a gate.
Keep your chosen AI in control of the conversation, and make the plan, approval, execution, and evidence visible as separate steps.