This paper introduces the "process harness," a new architectural mechanism designed to modernize legacy business workflows by integrating AI agents without replacing existing, reliable workflow engines. By placing a policy-governed agentic layer around a deterministic system, the authors enable organizations to add reasoning, adaptation, and oversight to their processes while maintaining the structural compliance and auditability of their original software.
The Process Harness Concept
Traditional business process management (BPM) systems are highly structured but rigid; they struggle to handle unanticipated situations, often forcing human workers to create "workarounds" outside the system. The process harness solves this by acting as an overlay. It intercepts specific control points—such as tasks, gateways, and sequence flows—to allow AI agents to intervene. Because the underlying workflow engine retains control over the process graph, the system ensures that the business process remains compliant with its original design while gaining the flexibility of an agentic system.
The Task-Decision-Flow (TDF) Model
To implement this, the authors developed the Task-Decision-Flow (TDF) model, which categorizes AI reasoning into three distinct agent types:
TaskAgents: Handle knowledge-intensive tasks, replacing fixed scripts with LLM-based execution.
DecisionAgents: Replace static gateway rules with per-case reasoning, allowing the system to make smarter routing decisions based on context.
FlowAgents: Monitor the overall process flow and use "hooks" to recommend structural changes, such as skipping, inserting, or removing tasks when necessary.
These agents operate within a "FRAME," an aggregate set of human-readable policies that define the boundaries of what each agent is permitted to do. This ensures that the AI's autonomy is "framed"—it is neither arbitrary nor opaque, but strictly governed by organizational rules.
Implementation in CUGA FLO
The authors present CUGA FLO as the practical realization of the TDF model. CUGA FLO functions as an orchestration layer that connects to an existing workflow engine. It uses a "Model Context Protocol" to bridge the gap between the deterministic execution of the workflow engine and the reasoning of the LLM agents. By demonstrating this on a loan approval workflow, the authors show how the system can handle both standard, frequent process variants and rare, complex cases that would typically require manual intervention, all without needing to rewrite the underlying process model.
Balancing Structure and Adaptability
The core contribution of this work is the reconciliation of two often-conflicting requirements: imperative requirements (the need for strict, predictable process execution) and normative requirements (the need for intelligent, policy-compliant adaptation). By separating these concerns, the process harness allows organizations to transform their legacy systems incrementally. This approach is reversible and modular, meaning that different levels of agentic autonomy can be activated or deactivated for specific processes depending on the need for human oversight versus machine-driven reasoning.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!