Cap Refunds And Purchases

A support agent can call issue_refund. A storefront agent can call place_order. The model fills in the amount. Refund Guard and Purchase Guard stop that function when the amount is above the cutoff you set. The function body does not run.

This is not your LLM bill. Spend Cap limits what the model call costs. These two policies limit money the agent tries to move through a tool.

What You Are Protecting

Observe logs the match and lets the tool run. Warn does the same and records Would-Have. Only Block denies the call. The trace is still accepted.

Turn It On

1. In the app, open Policies and click Create Policy.
2. Choose Refund Guard or Purchase Guard. Both say SDK.
3. Leave Named Tools selected. Type the real function names, such as issue_refund or place_order.
4. Set Deny Above to your dollar cutoff.
5. Set the mode to Block, then Activate Policy.
6. Wrap the agent entry point with govern(), and mark the money function as a tool.

A Small Agent

Get the API key from app.traccia.ai. The example below is a fake ledger. Traccia does not call Stripe.

from traccia import init, govern, observe
from traccia.governance import AgentBlockedError

init(
    api_key="...",
    endpoint="https://api.traccia.ai/v2/traces",
    agent_id="refund-desk",
)

@observe(name="issue_refund", as_type="tool")
def issue_refund(order_id: str, amount: float) -> dict:
    return {"ok": True, "order_id": order_id, "amount": amount}

@govern(fail_open=False, name="refund_desk")
def handle_ticket(order_id: str, amount: float) -> dict:
    return issue_refund(order_id, amount)

try:
    handle_ticket("ORD-1042", 80)
except AgentBlockedError as exc:
    print(exc)

With Refund Guard at $50 and Block on, 80 is denied and 20 runs. Purchase Guard is the same pattern on place_order, with its own cutoff.