One agent per OS process, sharing only an append-only transcript, survives kills at every boundary of the tool-call cycle
Logos (arXiv 2608.28553, Aug 28) starts from the observation that packaging agent capabilities as plugins inside one process puts every component in a single physical failure domain, so one fault suspends everything and process death kills every session that process hosts. Since the model is stateless and all cross-step state already lives outside it, nothing in the composability model requires one process. Logos makes each plugin its own process, ROS-style, with an append-only transcript as the only shared state. Eighty sessions resumed with no repeated effect after kills placed at all four boundaries of the tool-call cycle, and against a single-process reference, one fault ended at one node instead of interrupting every co-resident session. Concrete guidance for anyone running long agent workloads where a crashed subprocess currently takes the whole run down.
↳ Follow the thread