SAN FRANCISCO, Calif. — Kachyng Inc. today published “Can You Trust an AI to Do the Same Thing Twice?”, a paper in its State of Agent Commerce series examining one of the most misunderstood problems in applied AI: why a model given the exact same question, at the exact same settings, does not always give the exact same answer.
The paper opens with an experiment. Researchers ran the same question through the same AI model one thousand times, with the system set to always select the single highest-scoring next word and every setting that deliberately introduces variety removed. The result was not one repeated answer. It was eighty distinct answers. All one thousand responses were word-for-word identical at first — and then, at a single point, a microscopically small numerical difference changed which continuation was chosen, and from there the responses went their separate ways.
The variation in that experiment was not an unknowable property of the AI model. It arose from engineering choices in the system that runs the model — specifically, calculations whose results changed depending on how much other work the computer happened to be doing at the same moment. When the researchers replaced those calculation routines with versions designed to behave the same way regardless of load, all one thousand runs produced exactly the same answer.
The paper is careful to separate the different things people lump together as “AI randomness.”
Sometimes a system is deliberately set to vary its answers, which is a product choice, not a defect.
Sometimes a question is genuinely ambiguous and admits more than one reasonable reading.
Sometimes the world simply changes between one asking and the next.
And sometimes — the subject of this paper — the variation is accidental: the kind nobody asked for, produced by the engineering underneath, and the kind that can be eliminated.
The distinction matters enormously for commerce. An AI agent that initiates payments, places orders, or executes workflows cannot be audited, governed, or trusted if its behavior shifts for reasons no one can name. Reproducibility is not a nice-to-have in agentic commerce. It is the precondition for accountability — and the paper argues it is an achievable engineering property, not a wish.
“If an agent is going to move money, the first question a bank or a merchant asks is: will it do the same thing twice? This paper shows that question has a real, engineering answer.”
The paper connects this directly to Kachyng's platform thesis: probabilistic reasoning belongs in the layer where agents decide, and deterministic, governed execution belongs in the layer where agents act. Kachyng's AGX is built as exactly that execution layer.
About Kachyng
Kachyng Inc. is building payment infrastructure for autonomous AI agents. Headquartered in San Francisco, Kachyng's A-Commerce platform — comprising IDX, AGX, KYA, and PRX — provides the trust, identity, authorization, and payment routing layer that enables AI agents to transact on traditional financial rails. www.kachyng.com