OpenAI's Astra introduces "recurrent depth" reasoning—looping over a query internally rather than emitting sequential steps—which produces fewer legible chain-of-thought traces even though Astra's current use is limited and its CoT is still described as readable. If this technique scales, the CoT logs you currently rely on for debugging agent misbehavior, auditing rogue actions, and satisfying alignment/monitoring requirements will degrade or disappear, so build observability that doesn't depend on faithful chain-of-thought and watch whether Anthropic and DeepMind follow suit.