Parlant

Primary use cases

• building controlled customer-facing chat agents
• enforcing behavioral guidelines in conversation
• tool-using conversational workflows
• explainable agent decisions

Key concepts

• Guideline → deterministic-control-flow-not-prompt (docs) — A condition-action rule (for example, condition='customer expresses interest in a high-risk product', action='provide risk disclosure') that the engine matches only when its condition holds for the current turn, so behaviour is specified as discrete rules instead of one system prompt.
• Journey (docs) — A multi-turn SOP expressed as a state diagram the agent follows when guiding a customer through a process like a booking or a return; it is adaptive, so the agent may skip ahead, revisit states, or change pace based on how the customer responds.
• Attentive Reasoning Queries (ARQs) → attentive-reasoning-queries (docs) — Structured reasoning blueprints built into prompts that walk the model through targeted queries at the steps where guideline adherence tends to slip, used as distinct sets per engine component.
• Canned response (docs) — A pre-approved response template the engine can emit at high-stakes moments, capping failure severity by giving precise wording where it matters instead of letting the model phrase the reply freely.