An owner tells the service agent "this is about Bella, my golden retriever" in the first message. Three turns later, without repeating Bella's name, the owner asks "were her lab results back yet?" What configuration lets the agent carry the identified patient forward reliably instead of re-inferring it from context each turn?
Select an answer to reveal the explanation.
Short Explanation
A variable is the sticky note the agent writes Bella's name on the moment it learns it, then glances at for the rest of the visit. Without one, the agent is relying on the model's memory of the chat instead of a fact it actually stored.
Full Explanation
Session or agent variables let a value such as an identified PatientId be captured once and then referenced deterministically by every subsequent action in that session, which is the durable way to carry state like "which pet is this conversation about" across turns. Relying purely on an instruction asking the model to "remember" the pet is the closest-looking trap: it can work most of the time, but it depends on the model correctly re-deriving context from prior turns rather than reading a stored value, which is exactly the kind of non-determinism the exam expects you to design away. Re-running classification every turn is not how the loop works and would only re-select a topic, not track which patient is under discussion; classification and patient identity are different concerns. A guardrail that blocks actions until the name is restated actively breaks the desired experience by forcing the owner to repeat information the agent should already hold, and it is a behavioural constraint, not a state-tracking mechanism. Adjacent fact: variables are also how a specialist enforces deterministic personalisation, such as carrying a preferred contact method forward without re-asking.