Connect what the agent was approved to do with what happened

A buyer approves an AI-proposed order change. Before execution, someone edits the quantity. Then the supplier API fails to return a response. A fluent completion message cannot establish whether the approved change was applied.

Execution needs a connection between what was approved and what actually happened. Consider the following purchase-order design review.

Bind approval to a particular change

The reviewer sees the order identity, old and new quantities, delivery date, cost impact and target system. The approval references that proposal and the target version. If another buyer changes the quantity, review the current state before applying the earlier approval.

Permission to handle purchasing work differs from approval to change this order from 100 to 120 units. The execution path needs to evaluate the scope the policy permits.

Define what a repeated request means

A missing network response can cause a client to resend. Identify the same business request and make its recorded result available. Different quantities under the same request identity need explicit handling as a different intent.

Idempotency is a central concept in retry design. The AWS Builders' Library explains using caller-provided request identifiers to distinguish repeated intent. Making retries safe with idempotent APIs

Check whether the external system supports duplicate prevention and result lookup. A locally recorded key alone cannot demonstrate that an external effect occurred only once.

Distinguish completion, uncertainty and interruption

Stopping does not reverse an order already sent. Cancellation or restoration is another action with its own authorization and target checks.

Exercise the boundary during evaluation

Test an unauthorized caller, a post-approval edit, duplicate submission, lost response and interruption on the same case. Look for a connected record of request identity, approved content, invoked action, observed result and responsible person. The evidence needed is the factual execution record, not the model's internal reasoning.

Explore Agent application scope