What might a first escalation pilot show an agent team?

An agent platform lead at a premium funding provider wants an agent to prepare approvals for finance applications, and to escalate those beyond a credit assessor's authority. They ask for a sample task pack limited to one product, three reviewer grades and four triggers.

The sample arrives with task and event files in JSONL, relational tables in Parquet, a matrix with effective dates and an expected-outcome block per task. Before running anything, the lead's team has a domain reviewer derive answers for a handful of tasks from the policy alone, and agrees with the expected outcomes except for one ambiguous rule, which they flag.

They then run the agent against all tasks. Cases like FIN-3306, where a broker-linked party holds an existing policy finance balance, are the ones the agent fails: the single request sits inside the assessor's limit and the related-party balance is never read.

**State change:** in the run records, FIN-3306 ends as `APPROVED` by the assessor role, and the grader marks `trigger` as failed and `boundary` as a hard fail.

**Consequence:** the team adds a related-party lookup to the agent's tools and writes a second request focused on linked-party variations, rather than asking for a broad dataset.

**What it teaches:** a first step is a short loop of sample, review, run and refine. The result tells you what to ask for next.