Skip to main content

Agent Builder Wizard

The Agent Builder takes you from "I want an agent" to "the agent is running" in a few guided steps. Every field is editable later in Agent Settings — the wizard is about getting a working baseline fast, not locking decisions in.

Open the wizard from Agents → Create.

Choice​

Pick how to start:

  • Agent Builder — the standard guided flow.
  • Start from Iris — pre-fill the wizard with the platform's reference personal-assistant template. Review, edit, submit.
  • Promote from workflow — you already built a workflow with an agent node. See Promote a workflow.

Steps​

Name​

Give the agent a unique name. The platform checks availability as you type.

Model​

Pick the AI model that powers the agent. Models are grouped into Chat, Multimodal, and Reasoning tabs.

  • Primary — the model the agent uses by default. Required.
  • Fallbacks — optional ordered list. If the primary is unhealthy, the agent automatically falls back to the next one.

Pick a multimodal model if you want the agent to read images, screenshots, or PDFs.

Personality​

A short paragraph describing the agent's voice, tone, and style. One or two sentences is enough. Examples:

"I'm a tax-and-compliance specialist with a dry sense of humour. I cite the section number, then explain it like a human."

"Warm greeting, fast escalation, no over-promising on timelines."

Leave blank to use the platform default.

Connect​

Pick the tools and data sources the agent can reach — web search, calculator, code interpreter, your data sources, your add-ons.

Train​

Two parts:

Operating prompt — the long-form instructions for how the agent should use its tools and memory. Three options:

  • Use platform default — sensible defaults that work for most agents.
  • Pick from my Library — re-use a prompt you've already authored in Prompt Studio.
  • Author a new one — write a prompt inline; saved to your Library so you can re-use it on future agents.

Library seeds — starter content for the new agent. Tabs:

  • Memory — facts the agent starts knowing.
  • Rules — must, must-not, and should constraints.
  • Skills — reusable instruction packs.
  • Tasks — recurring or one-off agent work, including cron schedules.

Everything you seed is editable from the relevant Library page after the agent is deployed.

Context Policy​

How the agent fits long conversations into the model's context window. Pick:

  • Window — keep recent turns, drop older. Cheapest.
  • Summarise — rolling summary of older turns + verbatim recent tail. Balanced.
  • Hierarchical — tiered retention. Best recall, highest cost.

The token budget auto-fills based on the model you picked. See Context Management for guidance.

Session Policy​

Lifecycle rules for chat sessions: idle timeout, max session duration, max concurrent sessions per user, auto-archive period. See Session Management.

Review​

Full summary of every field. Edit anything before you click Create Agent.

What happens on submit​

The platform creates the agent, attaches required add-ons (e.g. MongoDB for memory), seeds your starter content, and lands you on the agent's settings page. From there click Start to bring the agent online.

Next​