The agent's entire personality and playbook, in plain language — who it is, how it greets people, what it should ask or offer, and what topics mean it should stop and call escalate_to_human instead of answering itself. This is the only place that controls what the agent actually says; there is no behavior hardcoded in its code.

An optional, usually cheap/fast first pass over the client's message. Its only job is to output a short category label (e.g. simple_faq, complex, needs_human) that the Routing table below uses to pick which model actually replies — so easy questions can go to a cheap model and hard ones to a stronger (and pricier) one. Leave it off to always use the Default model.

Maps a classifier category to the model that should handle it — e.g. send complex to a stronger model, leave simple_faq on something cheap. A category with no matching row here, or any message when the classifier is off, uses the Default model above instead.

An optional second model that reviews every draft reply before the client ever sees it. If it rejects a draft, its feedback is sent back to the agent to try again — if it keeps rejecting past the retry limit, the dialog is escalated to a human instead of ever sending an unreviewed reply. This is the main safety net against the agent going off-script.

Lets the agent call a search_knowledge_base tool mid-conversation, which asks the Knowledge Base's own agent (same RAG + judge pipeline as its test chat) and gets back an already-vetted answer to ground a reply in, instead of guessing or relying only on the Scenario text above.

Reviews every draft answer against the same retrieved passages before it's returned — rejects it if it states something the passages don't support. A rejected draft is sent back for a rewrite, up to the retry limit; after that, the last draft is still returned but flagged as not judge-approved.