The fix isn't a prompt asking the agent to be honest. It's wiring it so it has less room to invent: instead of asking it to recall a fact, you have it pull that fact from a real source, a database record, a document, an API response, a calendar entry, and answer only from what comes back. If nothing comes back, it needs to be able to say so instead of filling the gap with something plausible. We map the manual process before building anything, so we know which facts have a real source behind them and which don't, and we won't automate the second kind until there is one.
The second piece is a check before the agent treats anything as finished. It reads the situation, decides on an action, acts through a real tool, then checks the result against the source before moving on, retrying or handing off if the check fails instead of reporting success anyway. In a content engine that writes and publishes on its own, product facts come from the actual product feed, not from what the model half remembers about the catalog. In an operations system sending a report, the numbers come from a live query, not a paraphrase. We run this same pattern across our own operation, 110+ scheduled jobs and 40+ integrations, because it's the discipline that keeps our own systems from quietly making things up too.
None of this makes an agent infallible. A source can be wrong or out of date, and grounding only guarantees the agent isn't inventing on top of good data, not that the data itself is perfect. That's why the systems we build escalate to a person when confidence is low or a situation falls outside what was mapped, rather than pushing forward with a confident-sounding guess. If a task is mostly judgment calls with no reliable source to check against, an agent shouldn't be making that call alone yet.