Google Research describes Cogentic, a multi-agent harness where an orchestrator assigns proof directions to parallel provers, adversarial verifiers assume every step is wrong until justified, and a verified ledger accumulates confirmed lemmas across rounds. Notably, the orchestrator and advisor are explicitly forbidden from having mathematical opinions — no speculating whether a direction is promising. The claim is five new results on open problems at STOC/FOCS difficulty, using O(100) to O(1000) Gemini calls per problem, with no expert hints. The headline numbers: a Bulow–Klemperer two-sided market constant dropping from at least 20,000 per side to exactly 2 on the smaller side; a PoA of 1.5 for two-bidder autobidding with a matching lower bound; anytime regret matching the fixed-time constant; and a 5.2 → 3.52 revenue approximation improvement. The honest caveats are in Section 4: verification is by domain experts in the authors' own areas, not mechanical proof, and some companion papers include coauthors who were already working on the problems. The system also proved the PoA result the authors had conjectured but couldn't prove — including a part they hadn't studied. The gap they flag is real: it can generate candidate results faster than humans can read them. I ran a research group for years and the division of labor described — one partner who never does math, another who assumes everything is wrong, a third who summarizes history selectively — is just academia, industrialized. The innovation here is that nobody can quit for a faculty position.