Build your agent

Behavior rules & triggers

Behavior rules let the agent do more than chat — fire a CTA, capture a lead, route to a human — based on signals like keywords, URL patterns, scroll depth, idle time, or message count. Rules are evaluated server-side during prompt assembly and client-side for UI affordances.

Rule structure

Every rule has three essential pieces:

• Kind — one of cta, lead_capture, route, curated.
• Conditions — JSON describing when the rule activates: keywords, URL patterns, scroll depth, idle duration, message count.
• Action — JSON describing what to do when it fires.

Rules also carry an enabled flag and a priority integer. When multiple rules match the same turn, higher priority wins.

Visitor-side triggers

The widget watches a few lightweight signals locally — no extra request to the server:

• Scroll depth — fires past a percentage threshold.
• Idle — fires after N seconds without interaction.
• Exit intent — fires when the cursor moves toward the top of the viewport.
• URL match — pattern-based page detection.

Because these are local, response is instant.

Rule kinds

• CTA — renders a card with title, description, and one or two buttons (open URL, send message, capture lead, dismiss).
• Curated — bypasses RAG with a fixed answer, ideal for pricing, legal language, or support routing where consistency matters. See Curated answers.
• Lead capture — triggers an inline form after a conversation milestone (message count, low-confidence detection, intent signal).
• Route — escalates to a human operator when complexity indicators emerge.

Management & testing

Rules are edited at /app/agents/{id}/behavior with a kind-aware builder. The runtime allows up to 20 enabled rules per widget init — additional rules are silently skipped, so prune aggressively.

Variants can be A/B tested via Experiments and validated end-to-end in the Playground.