Anthropic just taught its coding agent how to be interrupted. Sounds small. Engineers spent 50 years solving this exact problem for normal software. Databases got transactions so a crash can't leave your data half-written. Unix got SIGTERM so a process can flush its buffers before dying. Even your laptop asks apps to save your work before shutting down. AI agents had none of that. Hit your usage limit mid-task and the thing just stopped, sometimes with a file half-edited, tests updated for…