When a tool goes silent after execution: retry or verify first?

The discussion under my previous post got me thinking about a more specific situation than simply saying "done": what if execution may have succeeded, but we cannot verify it?

Imagine an assistant sends a request to create a record, then the connection drops before the response arrives. The missing response alone does not tell the user whether the record was created. Retrying immediately could create a duplicate.

This is a hypothetical example for discussion, not a failure I have reproduced in Gemini.

For me, a useful message in that situation would be:

I sent the request, but I couldn't confirm the result. I'll check whether the record exists before retrying.

I'd like to hear from people building agents that use tools: how do you connect an execution attempt to a check of its outcome? How do you decide which operations are safe to repeat and which need verification or user input?

An error or success message alone isn't enough here. The user needs to know the next step without risking a duplicate operation.