Agent Harness: the architecture that separates AI demos from real planning systems

Today any model can answer interesting questions about supply chain. You can ask an AI to analyze geopolitical risks that may affect your supply chain, identify possible logistics disruptions, suggest sourcing alternatives or summarize the impact of tariffs and international conflicts on a specific industry.

That is useful. And, in many cases, impressive. But there is a huge difference between analyzing context and executing real operational planning.

An LLM can discuss stockout risk for a specific segment. It cannot, on its own, reliably recalculate a synchronized plan for demand, supply, inventory, capacity and production of a real multi-level operation. Much less run a corporate MRP respecting finite capacity, BOM explosion, lead times, operational constraints, inventory policies, shelf life, plant synchronization and the financial impact of the plan.

This is not a chat problem. It is an architecture problem. That is exactly where the Agent Harness concept comes in.

A large part of the market still treats AI as a sophisticated chatbot. You ask a question. The model answers. End of story.

That works for personal productivity. It does not work for Supply Chain Planning.

Industrial planning involves hard constraints, deterministic computations, complex business rules, measurable financial impact, multiple horizons and dependencies between modules. An AI agent cannot simply "invent" a production, inventory or capacity recommendation. It has to operate inside a reliable system.

The harness is the layer that surrounds the LLM. It is the set of structures, rules, tools, memory, validations and execution mechanisms that turn a probabilistic model into a reliable corporate operating system.

The LLM alone does not solve the problem. It is one piece inside the flow. Everything around it, allowed tools, accessible data, business rules, validations, orchestration across agents, persistent memory and governance, is defined by the harness.

The same model produces completely different results depending on the harness built around it.

In the context of Supply Chain Planning, the harness is five layers working together. The diagram below shows how a user request flows down through validation, orchestration and execution, then back up as a validated agent response.

The first layer protects the organization before any prompt ever reaches the model.

Input validation. Every request goes through a filter the company configures. The customer defines what cannot be sent to the LLM: tax IDs, personal identifiers, sensitive commercial data, intellectual property, anything covered by compliance. The filter is mostly deterministic, with explicit auditable rules, and can layer in lightweight semantic validation for ambiguous cases.

Compliance filter. Business rules that go beyond the technical layer. A pharmaceutical company can block formula data. A multi-country operation can restrict cross-plant data flows. The system honors those boundaries automatically.

Rate limiting. Usage control by user, team or module. Governs token consumption and prevents misuse.

Output sanitization. Validation does not stop at the input. Before any response reaches the user, it goes through a sanitization layer that checks consistency, removes unwanted artifacts and applies the same compliance rules in the opposite direction.

The orchestrator is the heart of the harness. It receives the user intent and decides which specialist agent to trigger, in which sequence, with which context.

At NPLAN there is no single generic agent. There is a coordinated network of agents specialized by domain: a demand agent, an inventory agent, a capacity agent, a supply agent, a finance agent. Each one operates with context limited to its domain, specific tools, clear objectives and its own rules.

The customer can configure the personality of each agent, adjust its behavior parameters or create fully custom agents for specific operational needs.

The orchestrator also handles error recovery. When an agent returns an inconsistent result, the system does not propagate the error: it tries to correct it, escalates to human review or informs the user clearly.

This layer solves a problem most AI solutions ignore: how to combine probabilistic reasoning with deterministic computation that has to be right the first time.

The full database never leaves the environment. In most interactions the agent receives only structural, obfuscated samples of the tables: schemas, representative examples and distributions, enough for the model to reason. When an answer requires a specific slice, part of that slice may travel to the LLM, always within the customer compliance boundaries. Even then, the full database stays local and every call is logged and auditable. It is a radically safer model than sending the whole database to the AI.

The supply chain engine. MRP, BOM explosion, finite capacity and inventory optimization run in a specialized math engine, deterministic and reproducible. The LLM does not compute: it triggers the engine and interprets the result.

Rich output. Agent answers are not just text. The harness turns results into HTML charts configured dynamically per context, planning grid widgets for structured data, and analysis-specific formats.

Next-step suggestions. At the end of every interaction the system proactively suggests the next relevant steps, keeping the conversation productive and contextualized.

LLMs have no memory across calls. The harness has to provide it intelligently.

Memory operates on three horizons: short term (the active conversation), mid term (parameters of the active planning scenario, horizon, policies, shop floor constraints, operational calendar, frozen periods) and long term (consolidated inventory policies, historical behavior, configured preferences).

The agent does not just receive conversation history. It receives live operational context: active scenario, frozen periods, capacity constraints, inventory policies, planning horizon, financial targets and operational priorities.

Specialized context engineering is what separates a generic agent from a supply chain agent. Sending a few lines in a prompt is not enough. The harness injects scenario parameters, strategic priorities, service targets, per-plant rules and financial indicators. That changes the quality of the decisions completely.

The more autonomy the agents gain, the more critical governance becomes.

Response validation. Every answer is validated before it reaches the user. The system checks internal consistency, verifies that the numbers make sense within plan constraints and applies deterministic verification loops. The agent does not just answer, it self-corrects continuously.

Query audit. Any reasoning that produced an answer can be inspected. If a hallucination is suspected, the analyst can audit the exact query that ran, the data used and the logic applied. Real traceability, not just a chat log.

Decision logs. Everything sent to the LLM is recorded. Decision history, versioning, full explainability.

Learning feedback loop. Every answer prompts the user to rate it as positive or negative. When negative, the feedback does not vanish into a forgotten dashboard, it feeds the harness improvement process. The system learns from mistakes in a structured way.

Horizontal platforms are commoditizing the basics fast: OAuth connectors, SaaS integrations, generic skills, agents that run administrative tasks.

But Supply Chain Planning is not a generic problem.

A pre-built workflow does not understand finite capacity, segment-based inventory policy, multi-level planning, BOM explosion, shop floor constraints, shelf life, variable lead time, plant synchronization or the financial impact of the plan.

It is impossible to solve this only with prompts.

Many people are still debating which model is best. That debate loses weight quickly because models are converging.

The real competitive edge is migrating to architecture, harness, context, governance, orchestration, integration with real math engines and vertical specialization.

That is exactly what separates an AI demo from a corporate operating system based on AI.

The future of planning will not be a chatbot screen replacing planners. It will be an intelligent layer working alongside specialized math engines, where planners still exist, but with much faster simulation, contextualized recommendations, automated exception analysis and domain-specific copilots.