System prompts hold up in demos and quietly fall apart once an agent talks to real customers; routed graphs turn brittle the moment a conversation drifts off the happy path. The wager here is that conversational control is really a context problem: at every turn, decide which rules and tools actually apply and surface only those, instead of stuffing everything into one prompt and hoping the model obeys.
What Sets It Apart
- Guidelines are condition-action rules re-evaluated each turn, so the agent stays inside business policy by construction rather than because a prompt happened to hold.
- Canned Responses swap free-form generation for pre-approved templates wherever a hallucination would be unacceptable — useful when a wrong sentence is a compliance incident.
- Journeys encode multi-turn SOPs that bend to how customers actually behave, instead of forcing them through a rigid flowchart.
- Explainability comes from full OpenTelemetry tracing: every reply is traceable to the specific rule that produced it, so audits stop being archaeology.
Where It Fits
Think of it as the opposite end of the spectrum from general agent frameworks. LangGraph targets workflow orchestration and DSPy optimizes prompts; this stays narrow, governing conversation turn-by-turn. That focus is the point — it trades open-ended flexibility for behavior you can predict and defend.
Who It's For
Great fit if you ship support, sales, onboarding, or advisory agents in regulated, high-stakes domains — finance, insurance, healthcare, telecom — where tone, accuracy, and auditability are non-negotiable; it already runs in production at banks. Look elsewhere if your real need is open-ended task automation or multi-step tool pipelines, where a workflow-first framework will fight you less.
