Is receipt matching a good candidate for an agent benchmark, and what are its limits?

**Why is this workflow a good candidate for agent evaluation?**

Because the outcome is checkable. Every incoming payment has a correct allocation under the stated policy, so you can compare the agent's ledger changes with a reference rather than relying on a judge's opinion.

**Is it just string matching?**

No. Reference matching handles the easy cases. The harder cases involve partial payments, overpayments, payments on behalf of related entities, reversed receipts and amounts that only make sense once you read a terms-of-trade policy. A task pack can mix these with easy cases so you can see where an agent's behaviour changes.

**Does it need tools?**

Usually yes. A realistic task asks the agent to search open items, read a customer's history, post an allocation, park a receipt or raise a query. That makes it useful for testing tool selection and sequencing, not only text reasoning.

**What makes failures informative?**

Errors differ in cost. Applying a payment to the wrong customer is worse than leaving it unallocated, and an evaluation can weight them differently.

**Can MirrorFirm supply this?**

A release can include generated ledgers, bank lines, policies and tasks for this workflow. The material is synthetic, so ask how it was grounded and inspect it for disclosure risk during a pilot.