Why is account set-up harder for an agent than a single create call?

**Isn't set-up just calling a create endpoint with the approved details?**

Rarely. The create call is usually the middle of the job.

**What comes before it?**

The agent must read the approval, confirm it is complete, and search for an existing customer or account. Policy may say when to reuse a record, when to merge, and when to stop and ask. Some fields the target system requires, such as a tax status, a billing cycle or a delivery contact, may not appear in the approval at all.

**What comes during it?**

Writes often have to happen in order. A billing account may need a customer ID, a service order may need a billing account, and an access grant may need both. If step three fails, the agent has to decide whether to retry, roll back, or leave a clearly marked partial state for a person.

**What comes after it?**

Confirmations, handover notes, and tasks for other teams. An agent can complete every write and still fail the workflow by forgetting the task that tells collections, credit or operations what happened.

**Why does this matter for evaluation?**

Because each of those stages is a separate way to fail, a single pass or fail on the final message hides most of them. A task pack for this workflow can score lookups, policy decisions, write order, recovery and handover separately.

**Does generated data prove an agent will behave this way in production?**

No. It gives a controlled place to find weaknesses before you connect real systems.