AI Agent Demo vs. Production Failure Analyzer

The fixture graph contains four named nodes and three connecting edges: intent parser, database query, data formatter and external action. The demo declares four completed steps while production declares two, failing at the third node. At seventy pixels per count, four completed steps measure two hundred eighty and two measure one hundred forty; the four total nodes also measure two hundred eighty. These are predefined browser fixtures, not measured production agents or an executed refund. The production label changes a local scenario selection. Demo cumulative tokens finish at sixteen hundred twenty. Production tokens jump from four hundred fifty to twelve thousand two hundred fifty after an assigned eleven thousand eight hundred token database payload, then reach thirteen thousand four hundred fifty. At one pixel per fifty tokens the initial bar is nine, database increment two hundred thirty six and final total two hundred sixty nine. Four hundred fifty plus eleven thousand eight hundred plus twelve hundred equals thirteen thousand four hundred fifty. The figure illustrates how an oversized payload can dominate context; the safety budget is not specified numerically and no real context overflow is tested. The demo latency label is one hundred forty milliseconds and production is four thousand eight hundred fifty with timeout. Their assigned difference is four thousand seven hundred ten. At one pixel per twenty milliseconds the bars measure seven, two hundred forty two point five and two hundred thirty five point five. Run trace advances a scripted timer every six hundred milliseconds, so these latency labels are not measured timer durations. Native export serializes the selected fixture even before completing a trace. Cytoscape is blocked offline; listeners bind before initialization, but reset and run then fail on the absent graph. Actual fixture export and original control behavior must be checked independently.