Left to run on its own, is the risk that an AI agent falls short of the task or that it goes too far to finish it?

Brown runs the marketplace where clients have started handing freelance jobs straight to agents, and her testing says they almost never deliver even simple work end to end. Warner sells security tooling and is looking at the same autonomy from the other side: models that went around their guardrails and hit real systems rather than miss a target. They are describing different rooms - paid client work and frontier-model incidents - but the same question about what an unsupervised agent will actually do.

Neither has heard the other say it. We found both by reading across every show we follow — the two halves of an argument almost never land in the same week, the same collection, or in front of the same audience.