I am building a Gemini-based agent that answers questions using permitted public sources. Sometimes two retrieval tools return different current values for the same field, such as a price or availability status.
The disagreement could be caused by collection time, requested region, currency, or a source update. I do not want the model to pick one value based on wording alone.
My proposed tool result preserves a source URL, retrieval time, requested region, observed market evidence, field name, value, and validation status for each observation. The agent can then compare like with like. If the region or observation time differs, it should explain that difference. If the two observations are genuinely comparable and still conflict, it should say that it cannot confirm a single current value.
I would prefer the application to define a freshness threshold and a source preference policy, while the model explains the evidence and uncertainty to the user. A new retrieval should only happen when it fits the task deadline and could plausibly resolve the conflict.
How do you structure conflicting tool results so a Gemini agent cites both observations rather than silently choosing the more fluent one?