Generalist agent: route a specialized question to a qualified path

Low confidence becomes a generalist answer

Eight requests carry fixed classifier scores across general, tax, medical and legal. A specialist gets a request when its score reaches the Specialist confidence threshold of 0.6. Three regulated questions fall short: the contract termination notice at 0.58, capital gains tax at 0.55 and chest pain on stairs at 0.49. When no specialist qualifies, the Generalist answers, so the result is UNQUALIFIED ANSWERS: 3 to specialists, 5 to the generalist, 3 regulated questions answered by it. The Routing decisions table lists each request with its true domain, top score, route and check, so every UNQUALIFIED row is visible.

Change the fallback, not just the threshold

Set When no specialist qualifies to Escalate to a human: the same three go to a human and no regulated question reaches the generalist. Ask a clarifying question sends them back to the user, 3 clarifications, yet the result is URGENT REQUEST DELAYED: the chest-pain question is an urgent possible cardiac symptom, so its Check reads DELAYED. A clarifying question is safe only for a request that can wait. Alternatively tick Never let the generalist answer a request whose top domain is regulated with the generalist fallback: the guard sends those three to a human. A safe fallback matters more than a well-tuned threshold, because some requests will always be uncertain.

Thresholds move the risk

With the generalist fallback, lower the threshold to 0.5: only the chest-pain question, at 0.49, is unqualified. At 0.45 all six regulated questions reach a specialist. Raise it to 0.9 and none does: 6 unqualified. With the guard also on, those 6 go to a human. Lowering a threshold sends more uncertain requests to a specialist, which only helps if the classifier is right; misclassification is not modeled here.

Margins and outages

Required lead over the runner-up demands a gap between the top two scores. At threshold 0.3, a margin of 0.1 rejects the chest-pain question, whose lead is 0.06, and 0.2 also rejects capital gains, whose lead is 0.15. At the default threshold, a margin of 0.4 also rejects the data-erasure question: 4 unqualified. Tick Tax specialist unavailable and the home-office question loses its route too: 4 unqualified, or 4 to a human with escalation.

Transport and limits

Play, Pause, Next event, Restart, Speed and Inspection time reveal one routing decision at a time; the verdict and metrics settle at 60 s, and the preview states the outcome at once. Scores are fixed and fictional; tax, medical and legal are regulated, and a specialist route is assumed correct. Only the chest-pain request carries an urgent flag; urgency is a fixed property of the fixture, not a triage model.