A Process Harness for Uplifting Legacy Workflows to Agentic BPM: Design and Realization in CUGA FLO

Traditional Business Process Management (BPM) systems provide rigorous structural guarantees but struggle with unanticipated situations, often forcing organizations to rely on manual workarounds. While Large Language Models (LLMs) offer powerful reasoning capabilities, integrating them into existing processes often results in either rigid automation (where the LLM is limited to a single task) or loss of structural compliance (where the LLM acts as an unconstrained planner). The authors propose a middle ground: an agentic process harness that preserves the deterministic execution of legacy engines while adding a layer of policy-governed autonomy.

A process harness acts as an overlay that intercepts control points in a workflow—such as tasks, gateways, and sequence flows—to engage LLM agents. The authors formalize this through the Task-Decision-Flow (TDF) model, which decomposes agentic reasoning into three distinct roles:

• TaskAgents: Handle knowledge-intensive activities, replacing static scripts with LLM-based execution.
• DecisionAgents: Replace static gateway routing rules with per-case reasoning based on process context.
• FlowAgents: Govern runtime structural adaptations (e.g., skipping or inserting tasks) via a hook mechanism.

Each agent operates within a 'FRAME'—an explicit, human-readable policy set that ensures agentic behavior remains compliant with organizational requirements. This architecture allows for 'framed autonomy,' where agents have the flexibility to handle the 'long tail' of process variants while the workflow engine maintains overall structural authority.

The authors present CUGA FLO as the practical implementation of the TDF model. CUGA FLO integrates with existing workflow engines, treating them as the 'imperative' backbone while providing a 'normative' agentic layer. By using a Model Context Protocol (MCP) bridge, the system ensures that agents are fully aware of the process model, execution state, and history at every step. The authors demonstrate this approach using a loan approval workflow, showing how the system can handle regulatory overrides and complex routing decisions without requiring programmatic changes to the underlying process model.