The failure mode is consistent enough to have a pattern: an agent operating on a multi-hour task loses coherence not at the start but at the handoff points — where sub-tasks complete, where tools return outputs, where earlier decisions should constrain later ones. A developer who spent weeks debugging one such agent found the logs clean and the context enormous; the agent still refetched data it had already retrieved and repeated steps it had already completed. The context window was not the constraint. The capacity to use it was.