Continuous Stateful Intelligence: Beyond the Turn-Based Agent

An agent’s commitments outlive individual requests. Continuous Stateful Intelligence (CSI) makes the agent’s lifetime the unit of execution: observations revise persistent state, reasoning advances through bounded model calls, and communication or action leaves the agent available to continue its work. We specify this architecture and evaluate how protected state, explicit revision, and retrieval preserve commitments across interaction boundaries. In 200 paired robot scenarios, persistent state achieves 93.0% physical success, compared with 49.5% after reset and 47.5% with a recent window, matching full replay. A four-key extension approaches replay success with approximately 1/46th of its visible memory. Language experiments identify a retrieval failure despite correct storage; a matched comparison favors exact lookup over the learned reader. During ongoing robot activity, scheduled corrections reliably update state, but we detect no improvement in physical completion.