The most dangerous property of a language model is not that it makes mistakes, it is that its mistakes read exactly like its successes. A fabricated statistic arrives in the same confident prose as a real one. For casual use that is an annoyance. For work you intend to act on, a market entry, a hiring plan, a code change, it is disqualifying.
The Agents platform treats this as an engineering problem with a structural answer: no deliverable reaches you until a separate agent, the Evaluator, has tried to knock it down.
A second agent whose job is doubt
When the Deep Research Agent cites a source, the Evaluator re-fetches that source and checks that it actually says what the report claims. When the Data Analyst computes a figure, the Evaluator recomputes it from the data. When the QA Agent reports a broken journey, the Evaluator re-probes the target. The Evaluator holds no loyalty to the team that produced the work, its only output is a scored judgment of the evidence.
Labels instead of confidence theater
Every claim in a finished deliverable carries one of three labels: observed, the agent saw it directly in a source, a dataset, or a live system; inferred, it follows from the evidence but was not seen directly; verified, the Evaluator independently confirmed it. What cannot be verified is said plainly, not dressed up.
This costs something. Verification passes take minutes and credits that a naive pipeline would skip. We think the trade is obvious: a report you can defend in a meeting is worth more than a faster one you have to caveat. The workforce is only useful if you can rely on what it hands back, so reliability is built into the structure, not promised in the marketing.