Skip to content
Zapito
Open App

AI Agent

The brain of your bot. It reads the customer’s message (plus conversation history), uses whatever is attached to its resource port, and produces a reply with a confidence score — or an explicit signal that a human should take over.

Ports

Section titled “Ports”

Config

Section titled “Config”
Field Status Behavior
Instructions Live The system prompt: who the agent is, its tone, its rules. The setup form writes this for you from your business name and description — edit it any time. Knowledge-base content is appended automatically.

There is deliberately no model picker and no temperature knob: the runtime uses one managed model configuration with a fixed, sensible sampling temperature for every agent. Decorative controls that the engine didn’t honor were removed rather than left in place.

The managed model

Section titled “The managed model”

Zapito provides and operates the model behind every agent. It’s chosen, tuned and — as better or cheaper models ship — swapped server-side, with no change on your end: your instructions, knowledge and flow stay exactly as they are. You never see model names, tokens or provider bills; your plan simply includes the model.

Agencies and advanced workspaces that must run on their own provider keys can ask about the flag-gated bring-your-own-key mode — it’s not part of setup and most customers should never need it.

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, "escalate": true|false}. The confidence value is what the Confidence Check branches on; both show in the Test chat trace (e.g. replied (confidence 0.92)).
  • The escalate signal: the agent sets escalate: true when a human should take over — the customer asks for a person, the attached knowledge doesn’t cover the question, or the request is beyond what it can do (refunds, complaints, account changes). An escalated turn goes straight to Human Handoff, regardless of any confidence threshold.
  • 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, your customer never sees an error — the turn degrades into the same human handoff, and the failure is recorded for you in the conversation.