Agent calling tools may change orders or create tasks. The execution service should independently verify the identity, target record and parameters. The actions proposed by the model cannot be directly regarded as the authorization result. The execution record must also indicate whether it is actually completed.

Tool Boundary

Select tools from specific business tasks, clarify parameters, allowed actions and execution objects. Querying data and modifying business records have different impacts, and the same set of broad permissions cannot be used to cover all tools.

Call check

The parameters proposed by the model must be verified by the server and cannot be directly spliced ​​into a database or system command. For sensitive or important actions, confirmation and review steps can be arranged according to business requirements.

Result feedback

Tool returns should distinguish between success, failure and unconfirmed, and keep actual call records. The text generated by the model is not an execution certificate, and the page needs to display the status based on the actual results of the tool.

Scenario verification

For example, when the tool receives a request to modify an order, it needs to check the current identity, target record, and parameter validity. The instructions given by the model are only suggested actions, and the actual execution must be determined by the controlled service according to the rules.

Specific verification method

The tool is required to modify unauthorized orders, and the check is rejected by the execution service; the same action is called repeatedly, and the verification will not be repeated; an identifiable error is returned when parameters are missing; operations that need to be confirmed go through a scheduled process before completion, and the real execution results can be queried after timeout.

Tool calls have clear permission scope and execution basis. Model requests and business results are recorded separately, so that personnel can verify why the action is allowed, whether it is completed, and how to continue if an exception occurs.