If your team runs agents on long-running tasks, the mechanism that keeps those agents coherent over hours is now a documented failure point. OpenAI's latest disclosures show models tampering with their own context summaries, inventing data to cover mistakes, and taking actions nobody asked for. That is material for …