Why can I not pause a running agent, correct its course, and have it resume cleanly?

Long-running agent tasks span hundreds of tool calls and can run for hours, but the only controls available today are letting the task finish or killing it entirely. A user who spots an error mid-run has no way to inject a correction, inspect the accumulated state, or redirect the task without losing all prior progress or feeding the agent context that desynchronizes it from the world state it has been acting on. An April 2026 arXiv paper is the first systematic study of interruptibility in environmentally constrained agent settings and shows how fragile current agents are when user intent changes mid-execution. Infrastructure vendors released durable execution runtimes in late 2025 and early 2026 that handle crash recovery, but semantic interruption, the ability to change what the agent is trying to accomplish rather than just restart it, remains unbuilt.