AI Agent
The brain of your bot. It reads the customer’s message (plus conversation history), uses whatever you’ve attached to its resource port, and produces a reply with a confidence score.
- In ← usually a trigger.
- Out → typically a Confidence Check or Send Reply.
- Knowledge & tools (bottom, dashed) ← Knowledge Base and Custom Tool nodes.
Config
Section titled “Config”| Field | Status | Behavior |
|---|---|---|
| Model | Live | Which model answers: GPT-4o, GPT-4o Mini, Claude Sonnet, Claude Haiku or Gemini 2.5 Flash. Defaults to Claude Haiku. Routed directly via your provider key — see Security & routing. |
| Instructions | Live | The system prompt: who the agent is, its tone, its rules. Knowledge-base content is appended to this automatically. |
There is deliberately no temperature knob: the runtime uses one fixed, sensible sampling temperature (0.3) for every model. Earlier builds showed a temperature slider and a “Smart routing” toggle that the runtime never read — both have been removed rather than left decorative.
Runtime behavior
Section titled “Runtime behavior”- The agent sees the last 20 turns of the conversation, so follow-up questions work.
- It must answer in a strict JSON contract —
{"reply": …, "confidence": 0..1}— and the confidence value is what the Confidence Check branches on. You’ll see both in the Test chat trace (e.g.model claude-haiku-4-5 replied (confidence 0.92)). - Attached Knowledge Base content is injected into the system prompt as grounding (see Knowledge Base for budgets); attached Custom Tools are described to the agent (see each node’s page for current limits).
- If the model call fails (missing key, provider outage), the error is sent as the reply and recorded in the conversation — nothing fails silently.