PatchOptic Secures Shared-State LLM Workflows with Verified Updates

Agentic workflows often involve large language models (LLMs) operating on a shared, structured state. Due to the inherent limitations of LLM context windows, models typically only receive a relevant fragment of the state for each step, a technique known as progressive disclosure. Current methods for creating these model-facing views, such as RAG or AST queries, address the read side but lack mechanisms to ensure the validity of local updates when applied back to the full global state. PatchOptic fills this critical gap by providing an optic-inspired interface for shared-state LLM workflows. It leverages projected reads and verified structured patches, where each workflow step explicitly declares its read view, authorized write region, and patch-source region. This design not only enforces runtime validity but also supports delegation, sub-workflow composition, and static reordering of independent steps. Benchmarking shows that projected reads reduce token cost and data leakage while maintaining output quality, and runtime verification effectively blocks contract violations.