Action checks vs. rules in the agent’s prompt.
The quickest way to limit what an agent does is to tell it: “never refund more than the order total”. That instruction is read by the model, and a model can be talked out of it. Audon checks the action itself, outside the model, before it reaches Stripe, Shopify or your system.
Side by side. Facts about Audon from the product and its docs.
| Aspect | Agent Action Control | Rules in the prompt |
|---|---|---|
| Where the rule runs | Outside the model, in deterministic code, on every action the agent proposes. | Inside the model’s instructions. |
| Fresh facts | Business rules read the provider first: captured amount, already refunded, refunds in flight, disputes, subscription state. | What the agent remembers or fetched itself. |
| Duplicates and retries | Same action, resource and amount within 72 hours is denied; an idempotency key never executes twice. | Depends on the agent noticing. |
| Approvals | A rule can hold an action for a person, with an expiry. | Not available. |
| Proof | Hash-chained evidence of each check, the decision and the verified outcome. | The conversation log. |
| Bypass | Holds only if the agent’s direct access to the same APIs is removed. | Holds only as long as the model follows it. |
When rules in the prompt are enough. Said plainly.
- The agent can only read, never move money, so a wrong answer is the worst case.
- Every action already goes through a person who approves it in another system.
Questions teams ask when they compare.
Do I still need instructions in the prompt?
Yes: they make the agent ask for the right things. Audon decides whether the action may run. A denial comes back with a code and a reason the agent can relay to the customer.
Does it work with any model or framework?
Yes. Audon controls the actions, not the model: the agent calls integrity.execute from the SDK or the Action Gateway over MCP, whatever model it runs on.
The products
- Agent Action Control: Checks every refund, credit, discount or subscription change an AI agent makes before it runs, verifies the outcome and keeps the evidence.
Last updated