Recoverability Framework for Maintaining State in Long-Horizon Agents

This research introduces recoverability as a system primitive to manage interruptions in multi-step agent tasks. The proposed reference architecture uses a behavioral contract to select validated starting points and permitted recovery actions, preventing the propagation of errors during task resumption.