Recently, I have been debugging several Agents that can run for 2 to 3 hours.
A very obvious feeling: Many problems are actually no longer at the model level.
As the task gets longer, what really starts to break down are: - session continuity - tool state - checkpoint - context drift.
Many demos look very strong.
But once you start entering long tasks, the runtime really exposes its complexity.
The Agent Runtime of 2026 is becoming more and more like a lightweight distributed system.