Overview
An AI Agent is a configurable, LLM-powered assistant with its own purpose, personality, and guardrails. Each agent points at a pre-configured model, defines how it should behave through prompts, and can be tested in an admin playground or reached by clients through a chat endpoint. For how Agents fit the platform, see Core concepts.Where to find it
AI Agents live under Orchestration → AI Agents.Properties
| Property | Type | Required | Description |
|---|---|---|---|
title | string | Yes | Display name. |
slug | string | Auto | Identifier — generated from the title if not provided. |
purpose | string | No | A short description (up to 100 characters). |
description | string | No | An extended description. |
provider | reference | Yes | The model the agent runs on (model + parameters). |
context_window | integer | No | How many past message pairs are sent with each call. Default 20, range 1–50. |
enabled | boolean | No | A disabled agent rejects playground and chat requests. Default on. |
Prompt & behaviour
An agent’s behaviour is assembled from several parts, in order:- System prompt — the agent’s main role and behaviour instructions.
- Guardrails — conditions under which the agent should stop or escalate.
- Brand guidance — tone, style, and voice.
- Additional instructions — extra runtime context.
- Contact data — when the conversation is tied to a contact, their name, email, phone, and attributes are appended automatically, so replies can be personalized.
Guardrails and brand guidance can be written inline or pulled from a
Content Instance — useful for reusing the same
policy or brand voice across agents. When an instance is referenced, it takes
priority over the inline text.
Models
An agent references a pre-configured model rather than raw API keys — a named bundle of an LLM model and its parameters (temperature, length, timeout). The available models are:| Provider | Model | Variants |
|---|---|---|
| Anthropic | Claude Sonnet 4.6 | Standard · Creative · Strict |
| OpenAI | GPT-4o | Standard · Creative · Strict |
| Gemini 2.0 Flash | Standard · Creative · Strict |
- Standard — balanced output.
- Creative — more varied and longer responses.
- Strict — more focused and deterministic, shorter responses.
Interaction modes
| Mode | For | Behaviour |
|---|---|---|
| Playground | Admin testing | A chat panel on the agent’s page. Stateless — each request carries the conversation so far; nothing is stored. |
| Chat endpoint | Clients / integrations | A public, API-key-protected endpoint. The conversation continues across turns via a session. |
Sessions
The chat endpoint keeps a conversation going with a session id:- On the first message, omit it — the server generates one and returns it.
- Send it back on each following message to continue the same conversation.
- Sessions are short-lived and expire after a period of inactivity; once expired, the next message starts a fresh conversation.
- History is trimmed to the agent’s
context_windowso it stays bounded.
Behaviour & rules
- Disabled agents reject requests — both the playground and the chat endpoint return an error if the agent (or its model) is disabled.
- Context window bounds memory — only the last
context_windowmessage pairs are sent to the model on each turn.
Tool use (letting an agent call functions or take actions) is planned but not
yet available in this release.
Seeds
AI Agents aren’t supported in Seeds yet.Governance & permissions
Only a super admin or Master can create, edit, enable/disable, and delete AI Agents.API access
Agents are managed through the User API: full CRUD, restore, a playground test call, and a list of available models. The client-facing chat endpoint is reached with an API key through the Contact API. See the API reference.Related
Conversations
Where agents hold ongoing conversations.
Content Instances
Source guardrails and brand guidance from content.
Events & Automation
Orchestrate when and how agents are used.