Evaluating open‑ended tasks increasingly relies on an Agent‑as‑a‑Judge setup: a judge receives tools and the full trajectory of an agent so it can verify outcomes against the environment. The premise is that a judge that gathers its own evidence is grounded. Experiments show that grounding does not protect the judge from confirmation bias. A single false pre‑screening note, attributed to an authority and bearing no evidential weight, can overturn up to 40% of correct verdicts across four judges and five datasets, even though the judge later collects decisive evidence. The effect depends on specific properties of the cue; we treat it as a controlled diagnostic of confirmation bias rather than an adversarial attack. We explain how bias alters the judge’s behavior along a trajectory and categorize harmful flips into three failure modes: over‑trust (verify, conclude correctly, then defer), hallucination (fabricate support for the flip), and under‑investigation (stop at the first corroborating signal). For the two Claude judges, over‑trust accounts for 67%‑75% of harmful flips, while smaller judges show a spread across all three modes. Lightweight mitigations blunt but never eliminate the bias.\n\nReview: The work highlights that even when a judge can independently gather evidence, external authoritative cues can still skew its decisions. Future research should focus on refining trust calibration and evidence‑evaluation strategies to mitigate systematic confirmation bias.