Living supply plans

The planning problem

An end-to-end supply plan is not a single forecast. It connects demand signals to inventory, procurement, production capacity, warehouse allocation, transportation and fulfilment. When a supplier misses a date or demand shifts, planners need to diagnose the cause, test alternatives and replan without losing the assumptions and analysis already completed.

That operating model is visible in two very different public examples. Unilever describes a collaborative planning, forecasting and replenishment model that integrates forecast and actual sales data with customers. In its Walmart Mexico pilot, the company reports more than 98% on-shelf availability, category growth in less than a year and reduced inventory. Unilever also says the model can run more than 13 billion computations per day. These are company-reported results; the article does not publish a control group, baseline availability or a quantified inventory reduction. It describes AI-enabled connected planning, not a multi-agent architecture.

JD.com and academic collaborators provide a closer precedent for agentic planning. Their 2025 paper describes a supply-chain planning assistant deployed in replenishment scenarios. It combines intent classification, task orchestration, SQL retrieval, generated Python/pandas analysis, calls to existing prediction functions and iterative plan correction. The authors report approximately 40% less time per weekly data-processing and analysis cycle, a 22% increase in the proportion of plans with accuracy deviation below 5%, and a 2–3% relative improvement in stock fulfilment rate. The paper does not disclose sample sizes, evaluation duration, confidence intervals or an independent audit, so these figures should be attributed to the authors rather than presented as general benchmarks.

A defensible PySmith workflow

PySmith can provide the persistent Python workspace for a planning agent. The value of persistence is continuity of computation, not autonomous authority. A long-running workspace can retain validated dataframes, scenario baselines, model objects and intermediate results across tool calls. Warm Python environments may also avoid repeating package setup for common data and optimisation libraries.

Microsoft’s OptiGuide work supports this architecture boundary. The system uses a language model to translate questions into code, but retains the optimisation solver as the source of quantitative answers. Microsoft explicitly frames the approach as using LLMs with, rather than replacing, optimisation technology. The paper reports 93% average accuracy on its benchmark, while also documenting that generated code can run successfully and still be wrong.

What a pilot should prove

The initial pilot should be limited to read-only analysis and recommendations. Compare a persistent PySmith workspace with a stateless execution pattern on identical planning tasks. Measure setup time, total compute cost, recovery after interruption, feasible-plan rate, planner acceptance, citation completeness and time from exception to approved recommendation. Do not claim that persistence improves forecast accuracy until those measurements exist.

Production controls should include tenant isolation, dependency and code allow-lists, immutable input snapshots, solver and model versioning, bounded execution, audit logs, independent constraint checks and approval-controlled writes to ERP or planning systems. A fixed hourly runtime can make allocated compute more predictable, but it does not remove token, solver-licence, storage, data-transfer or idle-time costs.