How the Agent Harness Operates in Slow Moving Inventory

A practical view of Ontology, Judgment, and governed learning in an operating supply-chain Agent

What the Agent does

The Slow-Moving Inventory Agent combines risk identification, root-cause analysis, and governed learning. Its Risk Score uses sales, purchasing, demand forecast, and BOM data to estimate how likely a Part Number is to move from Slow Moving to No Movement. Multiple statistical signals are combined with governed business rules to produce the assessment.

Its Root-Cause Analysis or RCA explains why the risk exists by classifying each flagged Part Number as Demand Decline, Forecast Error, Overbuying, Procurement Decision, Supplier Driven, Capacity Constraint, or Undetermined.

Together, these capabilities tell buyers and planners which Parts require attention, how significant the risk is, and what is causing it. The Agent automates repetitive analysis while preserving human judgment over the resulting business decision.

The three layers inside the Agent

1. Ontological Layer

The Ontological Layer gives the Agent a consistent understanding of the business. It connects each Part Number to inventory movements, historical sales, demand forecasts, purchase orders, suppliers, BOM relationships, and known future demand.

It also defines concepts such as Slow Moving, No Movement, Forecast Error, and Overbuying. The ontology establishes what these concepts mean and how they relate; it does not decide which category applies to a specific Part Number.

2. Judgment Layer

The Judgment Layer applies the Agent’s decision logic to that business context. It combines statistical signals, scoring logic, governed business rules, thresholds, RCA classification, and eligibility or exclusion criteria.

This layer determines how likely a Part Number is to move into No Movement, why the risk exists, and whether the Part should be included in the analysis. The results are presented through the dashboard, where buyers and planners review the evidence, make decisions, and provide feedback.

3. Agent Harness

The Agent Harness governs how operational feedback can improve the Judgment Layer. It does not allow an AI model or a user override to change production logic independently.

When a buyer or planner disagrees with an RCA classification, risk assessment, or eligibility decision, the Harness links that feedback to the original result, the reason for the correction, and the business context available at the time.

The governed learning cycle

• Capture: Records the disagreement, supporting context, and user explanation.
• Explain: Organizes the evidence and identifies similar cases. An LLM can support this analysis but cannot change production logic.
• Propose: Transforms a supported pattern into a candidate rule, scoring adjustment, threshold change, or exclusion.
• Test: Evaluates the candidate against historical snapshots and compares it with the current production baseline.
• Govern: Reviews the evidence against defined business and performance thresholds and requires explicit human approval.
• Deploy and monitor: Releases only the approved version through a traceable process with production monitoring and rollback capability.

Operational feedback becomes evidence. Evidence becomes a candidate improvement. Only tested and explicitly approved changes become part of production decision logic.

How the layers work together

Suppose buyers repeatedly change an RCA classification from Forecast Error to Overbuying. The Ontological Layer provides the inventory, forecast, purchasing, and demand context and defines both categories. The Judgment Layer applies the current rules and produces the original classification. The dashboard allows the buyer to correct the result and explain why.

The Agent Harness identifies comparable cases and evaluates whether the feedback represents a recurring pattern. A candidate improvement is then backtested, compared with the production baseline, and reviewed. Only after it passes the required tests and receives explicit approval can it become part of the production Judgment Layer.

The same process can improve a risk threshold, scoring signal, or eligibility rule. For example, Parts held against confirmed future demand may require an exclusion rather than a different RCA classification.

From prediction to governed learning

• Ontological Layer: establishes the business context and common definitions.
• Judgment Layer: determines the risk, its cause, and the applicable decision logic.
• Dashboard: enables human review, action, and structured feedback.
• Agent Harness: turns feedback into tested, approved, versioned, and traceable improvements.

The Agent does not learn by silently changing itself. It continuously incorporates validated business knowledge over time while preserving human oversight, traceability, reliability, and business control.

Companion article: The Harness Behind Agentic Learning sets out the governed learning loop and the four guardrails behind every production change.