Can a tool call run without a user approving it first?
Who runs the function
The model returns a request to call a tool, but your application runs it. That means you can add checks, logging or a human approval step before any function executes. Treat the model's request as untrusted input, the same way you would treat a form submission. This design also gives you a clear audit trail.
Read-only tools, such as looking up a public price, usually carry less risk. Tools that send emails, delete records or move money deserve a stronger gate.
Where confirmation fits
A confirmation step can show the user the exact action and its arguments before running it. The user approves or rejects, and your code logs the decision. This keeps the model from acting on a misunderstanding or a manipulated prompt. Make the approval message short enough that people actually read it.
Set limits too, such as maximum amounts, allowed recipients or a rate cap. Those rules protect you even when approval is automatic.
- Require approval for changes that cannot be undone.
- Show the exact arguments to the user.
- Log every requested and executed call.
- Set limits on amounts, rates and targets.
Common mistakes
- Letting the model's request run without any validation on its arguments.
- Assuming approval prompts alone protect you without limits on what the tool can do.
