The Exception Queue

Why this game.This is what the current generation of planning software does: an agent plans everything, people manage by exception, and the planner's job becomes keeping the plan's inputs true. When I built agents like this one, the hardest question was never the forecast. It was what to put in front of a person. This page lets you sit in that seat.

The company.Merrimack Supply Co. is fictional: an HVAC and building-products distributor in New England. Three distribution centers, thirty-six branches, eight categories from air filters to compressors, ninety-six items. Demand follows each category's season with random noise, and the noise is fixed by the quarter's seed, so two players on the same seed see the same sales.

The agent.A replenishment engine of the kind planning vendors ship: a forecast per item per site (exponential smoothing on smooth items, a separate method for lumpy ones), order-up-to targets by service segment, time-phased requirements at the warehouses with supplier lead times, minimum orders and case packs, and an expedite when a critical item is about to run a warehouse dry. Every move gets a risk score from how big it is, how strange the sales signal was, its dollar value, its type and its item's importance.

The gate. The agent acts alone on any move that scores below your dial. The rules on top of the dial are yours: always review expedites, always review A items, never bother with small C-item moves.

The inbox. Promotions, new and lost accounts, supplier slips, dead sales feeds, cold snaps, a warehouse running at half speed. Some arrive early, some late, some never. Entering one puts it into the plan: the agent forecasts with the uplift, books estimated sales through a dead feed, and shows revised dates on late orders. Some messages only sound like events.

The models. Four language models, each on the job it is best at, routed by the server and named on the page whenever one does something. A cheap, fast model reads the inbox and proposes entries. A reasoning model with tools that run in your browser works a case with you, runs the engine forward to test a number, and hands you a one-click recommendation. A third ranks the whole queue into a morning brief. A fourth writes the debrief. None of them execute anything; every change is your click. The numbers underneath come from statistics and inventory policy, the way real planning systems keep the arithmetic away from the language model.

The desk. A budget of minutes a week. An entry costs twenty, a review costs four to fifteen depending on the case, a phone call costs five and sometimes turns up something nobody has told you yet. Anything you do not reach is stamped as the agent proposed it.

The score. The cost of the quarter is lost margin, stock carrying cost, expedite premiums and transfer costs. Your desk is placed between two references run on the same quarter: the agent alone with nothing entered, and the best reference desk. The share of the gap you closed is your score. Every reference is a real run of the same engine, computed in your browser after you finish.

Where it runs.The engine runs in your browser. The model calls go through this site's server to the providers with the message text, the case and the numbers they need, and nothing else. Your best score is kept in your own browser. Calls are rate-limited per visitor and capped per day.