Choosing an agent shape
Before you start
Agents need an AI model to think and a connector to act. Make sure your workspace has a connected LLM provider and, for anything that touches your data, a connected NetSuite account.
Build an Assistant step by step
An Assistant is a chat-first agent — a domain expert that plans, retrieves, and acts, looping you in when it needs a decision. It is created with a Chat Trigger and a default graph of Chat Trigger → AI Respond, so you have a working agent the moment you create it.1
Open Agents and start a new agent
Go to Agents in the workspace, then click New agent. The Create a new agent dialog opens with the three shapes.
2
Pick the Assistant shape
Choose Assistant Agent (“Domain expert”). Erstan creates it immediately with a default name (for example, New Assistant) and drops you into the visual builder.
Workforce (“Orchestrator”) and Automation (“Set process”) are the other two shapes. A Workforce routes work across other agents; an Automation runs a multi-step process from a Form or Schedule. See Agents overview for when to use each.
3
Name and describe it
In the builder, give the agent a clear name and short description (for example, AR Aging Assistant — summarizes overdue invoices and customer balances). This is what teammates see in the Agents library.
4
Confirm the trigger
Your Assistant already has a Chat Trigger, which means people start it by chatting or
@-mentioning it. You don’t need to add another trigger for a chat-first agent.Want it to run from a form or on a schedule instead? That’s a different shape — see Triggers.
5
Add steps (optional)
The default AI Respond step is enough for a capable assistant. To add behavior, open the node library on the canvas and drag in steps from Triggers, AI Steps, By Connector, or Utility, then connect them with edges. See The visual builder and the Step reference.
6
Set instructions, skills, and tools
Use the inspector to tell the agent how to behave:
- Instructions — plain-language guidance (“Always show invoice number, customer, amount, and days overdue”).
- Skills — reusable instruction packages it should load. See Giving agents tools, skills & connections.
- Tools / connections — the connector actions it may use, such as NetSuite reads. A step that needs an unconnected system shows a Connection required banner until you connect it.
7
Preview before publishing
Click Preview in the builder header to run the agent without making it Live. Provide a chat message, watch the run stream live in the Steps panel, then iterate. Preview runs are tagged Preview run in run history.
8
Publish
When it works, click Publish. The agent’s status becomes Live and it appears in the library for your team to run.
For triggered agents (Form, Schedule, Webhook, Email), publishing is gated until the trigger’s Run As identity is configured and consented. A plain chat Assistant has nothing extra to configure. See Preview, publish & versions.
The visual agent builder
Let the AI builder do the work
Inside the builder you can describe what you want in plain language and have Erstan edit the agent for you — adding steps, wiring connections, and writing instructions. Use it to scaffold a first draft, then refine by hand. See The visual builder for details.After you publish
A Live agent can be run on demand, assigned as a task, or started by@-mention in chat. Every run produces a tracked trace with live status and approvals.
You now have a Live Assistant. Run it from the library or the agent’s runs page, and watch its trace in Run history.
Next steps
The visual builder
Master the canvas, the node library, and the inspector.
Running agents
Run on demand, assign as a task, or schedule recurring runs.
Giving agents tools & skills
Attach the skills, tools, and connectors your agent needs.
Explore the Marketplace
Install a ready-made agent instead of building one.