Describe the bug
The new tool_confirmation feature that's supposed to facilitate human-in-the-loop agentic workflow is very inconsistent.

Suppose i have a tool that does entity extraction and I would like the user to confirm the correctness before proceeding with subsequent steps:

FunctionTool(entity_extraction, require_confirmation=True)

Sometimes it chooses not to wait for the confirmation and just starts to execute the next steps

Good example:

Bad example:

This is similar in nature with #1851 but it's closed without resolving it completely.