Latent reasoning could turn AI agents into harder-to-monitor black boxes
New reasoning designs could let AI agents think orders of magnitude longer without putting that work into human-readable chain of thought, making today’s strongest practical monitoring tool much less useful. Redwood points to COCONUT, which replaces CoT, and full-bandwidth transformers, which add a hidden channel alongside it; future agents could also communicate without understandable language. That would make agent swarms harder to inspect and could increase takeover risk.