Long‑horizon autonomous intelligent systems are built from heterogeneous components such as large language models, databases, external APIs, and rule engines, and their external state changes continuously during execution. Re‑executing the entire workflow after each change leads to massive redundant computation. This paper introduces an incremental consistency execution approach that relies on task‑fact contracts, field‑level dependency masks, and invariant domains for state perturbation results. After an initial verified run, conservative invariant domains are created for critical inputs; these domains are then used to decide whether downstream results can be safely refreshed without invoking expensive components again. When re‑execution is necessary, only the minimally affected output fields are recomputed, and an equivalence barrier blocks unnecessary downstream propagation. A version‑consistency gate at submission time further guarantees the safety of side‑effecting actions. Experiments on industrial fault diagnosis, enterprise analytics, and LLM‑based multi‑tool assistants demonstrate a substantial reduction in costly component calls and end‑to‑end latency while preserving high consistency and low incorrect‑reuse rates.
Review