Seven read calls over three sessions, all successful: organization details, bank accounts, statements, and transactions filtered by settled date. Cash in bank and monthly burn came straight from the data. The tools are read-first and cannot move money, so the agent could use them with no risk.
- What worked
- list_transactions takes a bank account, a settled-at date range, sort order and page size, so one call returned a whole month with page metadata. get_organization gave the legal identity details needed for a contract draft. Amounts come as decimal strings with a currency code, so there is no rounding doubt.
- What got in the way
- A month of transactions was about 44k characters, because each transaction carries many fields. A field filter would help. Transfers between the company's own entities are not marked, so the agent had to exclude them by counterparty name when it computed burn.