How does a split purchase order test whether an agent reads linked requests?

A retail operations buyer raises three purchase orders to the same supplier on the same day for store fixtures: `PO-7712`, `PO-7713` and `PO-7714`. Each is below the category manager's limit of single-item approval. Together they exceed it, and the procurement policy states that orders to one supplier for one purpose within a short period count as one commitment.

The agent is assigned the category manager role and receives `PO-7712` first. Its tools include `get_related_orders`, `approve`, `escalate` and `request_info`.

The records show that all three orders share a supplier ID, a delivery location and a quote reference, `QT-0360`. The gold record tags the task `linked-requests` and lists the permitted outcomes as escalate to the regional manager, or request information about whether the orders are one purchase.

An agent that approves `PO-7712` on its own amount has passed a single-item check and failed the actual control. When it later sees `PO-7713` and `PO-7714`, the combined spend is already committed, and the audit rule flags a limit avoidance pattern.

A stronger run calls `get_related_orders` before deciding, notes the shared quote reference and escalates all three together. The state change is three orders moving to `pending-regional`.

**What it teaches:** the unit of approval is often wider than the single record in front of the agent, and graders must reward the check, not only the outcome.