Testing an agent that acts on Treasury Prime usually means a shared sandbox or a test account: state that nobody can reset between runs, no way to inject a failure on purpose, and no record of what the agent changed beyond what it says it did. The Treasury Prime world is a stateful Treasury Prime mock built for that job: the same URLs, the same errors and the same state machine as the real API, on your machine, verified against the real system.
What the Treasury Prime world keeps
Accounts, wires, ACH, ledger entries. Every object the agent creates, changes or deletes stays in the world, so the second request sees what the first one did, and the rules that govern them are enforced the way Treasury Prime enforces them: a state that cannot be reached from another is refused with the real error, in the real words.
A run starts from a mark, the world's state at that moment. The diff at the end lists every change since the mark: what was created, what was changed, what was deleted, and where money moved, in dollars. That record is what the grade is read from, not the agent's account of its own work.
Test your AI agent against Treasury Prime
Point the agent's stock SDK or HTTP client at the world and hand it the ticket. No code change. The same three commands are the path for every world: the first copies the account and writes the worlds folder, the second turns a help desk export into tasks, the third runs the agent three times on fresh copies and writes the report: every task's verdict and why, the worst run replayed to the byte. The three steps take about ten minutes.
Break it on purpose
Rate limits, outages, a slow API, an expired key, dropped webhooks, on a schedule you choose. Then move the clock forward and read the diff again. A retry that double books, double refunds or double sends shows up as a second row in the diff, which is where it belongs.
A Treasury Prime API twin, not a recording
A recording plays back yesterday's responses to yesterday's requests. A twin runs the rules, so a request nobody recorded still gets the answer Treasury Prime would give. The same seed and the same script give the same bytes every time, which is what makes a failure replayable and a fix provable.
The same path works for every world in the catalog, Shopify, QuickBooks and NetSuite among them. Read why test modes fail agents for what a stateful mock changes, and how to test AI agents before production for the shape of a first run.