Workflow as Knowledge: Semantic Persistence for LLM... | AI Research

Workflow as Knowledge: Semantic Persistence for LLM-Mediated Workflows proposes a new way to handle complex AI workflows. Currently, most LLM applications treat workflows as temporary sequences of code or chat logs that disappear once a task is finished. This paper argues that we should instead treat these workflows as permanent, inspectable "knowledge objects." By doing so, developers can ensure that the reasoning, decisions, and intermediate steps of an AI agent are saved, allowing them to be reviewed, resumed, or audited long after the initial execution.

A New Conceptual Model

The authors suggest moving away from viewing workflows merely as executable scripts. Instead, they propose a model where every part of a workflow—the definition, the running instance, the context, and the final records—is stored as a persistent object in a shared knowledge substrate. Inspired by Lisp-style programming, this approach treats the workflow as a "live image" where the structure of the process and the data it produces are equally important. This allows users to look back at a workflow and see not just the final output, but the entire history of how that result was reached.

Distinguishing Computation from Judgment

A central feature of this model is the clear separation between two types of operations: derive and infer.

• Derive refers to deterministic, rule-based computation. These are standard calculations or data transformations that are predictable and repeatable.

• Infer refers to judgments made by an LLM. Because these are mediated by the model, the system requires that the context used for the judgment be explicitly declared and saved.
  By forcing this distinction, the model ensures that developers know exactly when a process is relying on a rigid rule versus an AI-generated decision. This makes the system more transparent and easier to debug.

Persistent Records and Human Oversight

The model emphasizes that human interactions—such as approvals or structured deliberations—should not be treated as fleeting UI events. Instead, they are captured as typed records within the knowledge substrate. For example, a "panel" record would store the motion, the arguments presented, and the final decision. This ensures that the reasoning behind a human-in-the-loop decision is preserved alongside the machine-generated data, creating a complete audit trail of the entire process.

Limitations and Future Directions

It is important to note that this paper is a conceptual proposal rather than an implementation guide or an empirical study. The authors emphasize that while they have defined the vocabulary and the object model, they have not yet built a formal system or tested it in a production environment. Furthermore, they clarify that simply saving these objects does not automatically guarantee that a system is trustworthy or reproducible; it merely provides the necessary structure to make those qualities possible. Formal transition semantics and empirical evaluation remain tasks for future research.