Why does Burna AI use twelve agents instead of one model?

Because each decision in grading and attribution is easier to constrain, test and explain when one agent makes it. Term extraction does not also judge causality; causal attribution does not also calibrate confidence; confidence calibration does not also encode the regulatory output. Each agent's output narrows what the next may produce, so the space of possible answers shrinks at every step and a result that contradicts an upstream finding is not passed on. The practical effect for a reviewer: every step leaves evidence behind, so an investigator, a monitor or an inspector can see why a grade was proposed, not only what it was.