- The paper proposes a conceptual object model that stores workflow definitions, resumable instances, inference records, approvals, panels, context snapshots, and dependencies as queryable knowledge objects.
- It separates deterministic, replayable derive operations from mediated infer operations, preserving executor control while making model context, influence, validation, and authority inspectable.
- An exploratory scan of 77 workflow artifacts informed the vocabulary, but the paper provides no implementation or empirical validation and identifies formal semantics, lifecycle policy, and human evaluation as next steps.
Motivation and problem statement
LLM applications have converged on explicit workflow structures—typed steps, tool calls, checkpointing, human approval gates, and resumable state—as a remedy for the opacity of unconstrained reactive prompting (Wang et al., 14 Apr 2026, Khattab et al., 2023, Fan et al., 2024). The paper "Workflow as Knowledge: Semantic Persistence for LLM-Mediated Workflows" argues that this shift solves only part of the representational problem. Even when control flow is explicit, the artifacts of execution remain fragmented: definitions live in source code or configuration, running occurrences in runtime layers, and model outputs, approvals, and deliberations in logs, traces, or chat history. These artifacts can be correlated by infrastructure, but correlation is not semantic role assignment.
The paper's central proposal is semantic persistence: workflow definitions, workflow instances, inference records, context snapshots, and dependency relations should persist as typed, queryable knowledge objects in a shared knowledge substrate—the same substrate that holds the knowledge those workflows produce. The distinction drawn is between execution persistence (checkpoints, logs, traces retained by systems such as LangGraph or AgentSPEX) and semantic persistence (workflow objects retaining declared roles as first-class knowledge after execution). The authors are explicit that the contribution is conceptual: no implementation or empirical evaluation is provided, and formal transition semantics are deferred to future work.
Architectural framing
The model separates three layers. A lower runtime service layer supplies model adapters, tools, external processes, and persistence facilities. A middle control layer contains the DSL machine—an executor that interprets declared objects, assembles context, mediates model and tool calls, validates results, applies transitions, and exposes the substrate interface. A higher semantic layer holds workflow definitions, instances, and their linked inference, approval, and panel records. The "agent harness" vocabulary from practitioner literature is treated as packaging of the middle and lower layers, not as a definition of what a workflow is.
The Lisp tradition supplies an explanatory lens rather than an implementation commitment: symbolic forms close to the structures they denote (following McCarthy's s-expression tradition), object identity, reflective inspection, and Symbolics Genera's live-image style of integrated, inspectable environments. The authors carefully qualify this analogy—they do not claim Genera addressed LLM workflow persistence; they borrow the intuition that programs, data, and user-visible state can inhabit one inspectable environment.
The derive/infer boundary
The most consequential semantic commitment is the distinction between derive and infer. derive denotes deterministic computation over available workflow state—testable and replayable. infer denotes mediated LLM judgment requiring declared context scope, prompt intent, expected return type, validation, result persistence, and capability policy. Non-LLM nondeterminism (randomized algorithms, stochastic simulations) is classified as derive when seed, inputs, version, and replay policy are explicit; the authors concede that when these cannot be made explicit, a separate primitive may be needed.
Two boundary conditions deserve emphasis. First, the split is explicitly not a security proof: an LLM may fill values at declared inference sites, and those values may influence later branches, but the executor retains all transition authority. When an inference influences a branch, the instance records a dependency relation from the inference record to the executor-applied branch—the LLM is a "mediated peripheral," never a process supervisor. Second, the authors acknowledge the boundary requires authorial discipline: some tasks ("find the next unresolved sub-question") could be implemented either way, and the DSL's job is to make the choice visible rather than pretend it is mathematically obvious.
Object schema and human-facing records
The core schema distinguishes active state (execution memory) from durable records (evidence), and within human-facing records separates an approval (an authorization event permitting, rejecting, or deferring a transition) from a panel (a structured deliberation occurrence with motion, options, arguments, decision, and context). This separation emerged during synthesis of the exploratory scan; notably, approval was not in the original scoring checklist and therefore has no corpus count—a candid limitation of the pilot evidence.
Every infer call carries an inspectable context snapshot recording the bounded view shown to the model service, aligning with practitioner accounts of context management as a first-class engineering concern. Workflow instances persist as resumable objects supporting narrow, policy-gated reuse of prior records (e.g., reusing a compatible panel-record instead of repeating deliberation)—explicitly scoped reuse, not a claim that repeated interactions are interchangeable.
Relation to provenance and existing systems
The positioning against related work is careful. Provenance systems (W3C PROV-DM, decision provenance, PROV-AGENT) capture histories largely after the fact; the proposed model assigns DSL-level semantic roles—infer, approval, panel—before any provenance export. The model is described as provenance-compatible but not provenance-complete: Appendix C maps objects to PROV analogues while identifying additions PROV-DM does not define (executor mediation, infer boundaries, approval authority, panel structure). Objects without direct PROV analogues include active state, guard, and executor-applied branch transitions whose influence is recorded without granting the LLM control-flow authority.
Against agent workflow systems (AgentSPEX, LangGraph, DSPy, Flows/aiFlows, WorkflowLLM), the contrast is ontological rather than syntactic: those systems solve the execution-control problem; this paper addresses the knowledge-persistence problem. Against Lisp-for-LLM work (persistent metaprogramming loops, Pel), the design is more conservative about authority—the LLM does not generate or evolve executable expressions autonomously.
Evidence status
Section 4 reports an exploratory scan of 77 selected skill-like workflow artifacts (17 local/internal, 30 Codex/Claude-style agent artifacts, 30 operational workflow artifacts), scored qualitatively on a 0–3 checklist. The authors repeatedly caution that this is a vocabulary-design probe, not empirical validation: the corpus was selected rather than random, counts describe only that corpus, and local/internal material is not public evidence. Findings include high prevalence of resource, guard, step, context, and record patterns across groups; higher scores for state, tool, resume, and serialize in operational artifacts, suggesting side-effecting capabilities and recovery behavior need explicit modeling; and identified refinements (capability/action, handoff/promotion, four distinct forms of context).
Limitations and open questions
The paper is unusually forthright about its status. It concedes that semantic persistence does not imply accurate attribution, reproducibility, trust, or audit quality, and that whether the affordances improve audit quality or trust is an open empirical question. No record-lifecycle policy is chosen: retention, compression, supersession, deletion, and tombstoning rules—and what provenance survives each operation—are unresolved governance questions. The derive/infer boundary lacks formal semantics; Appendix A is explicitly a sketch, not a calculus. The panel construct has no direct transition authority, but its interpretation, contestability, and credit assignment remain unevaluated. Storage and indexing costs of durable inference and context records are acknowledged but unmeasured. Whether the primitive vocabulary generalizes beyond skill-like artifacts to runbooks, research protocols, and operational procedures is untested.
Conclusion
This paper contributes a vocabulary and object model—not a system—for treating workflows themselves as persistent knowledge. Its four commitments are that definitions be data objects, instances persist as resumable checkpoints, derive and infer be held to distinct semantic standards, and human reasoning interactions persist as typed records with context snapshots. The value of the account lies in making authority and influence relations inspectable before execution begins, complementing rather than replacing provenance vocabularies and checkpoint-based engines. Its claims await exactly what the authors specify: formal transition semantics, lifecycle policy, a small prototype compared against a checkpoint-and-trace baseline, and human evaluation of attribution, review quality, and reproducibility.
