An AI agent can be right today and wrong tomorrow.

And it all depends on the prompt, We have to ask the agent to pull the latest from the web and use them to perform the calculation, otherwise the agent uses the outdated rules.

Better would have an evaluation mechanism which we should use against the ordinary math code for comparison and flag any mismatch