Step reference

When you build an agent in the visual builder, you assemble it from steps (also called nodes) on the canvas. The node library groups them into four categories: Triggers, AI Steps, By Connector (actions), and Utility. This page is a reference to what each step does so you can pick the right one.

Every agent needs exactly one trigger to start it, and most pipeline agents end with an End step. AI steps and actions go in between.

If you are choosing between AI steps rather than looking up a specific step, start with Choosing AI steps.

Triggers

A trigger decides how an agent starts. Add one from the Triggers group; it becomes the start of the agent. See Triggers for setup details, including Run As for automated triggers.

Step	Starts the agent when…
Chat Trigger	Someone sends a message in chat (Chat message). Forces the agent into a chat-first shape.
Form Input Trigger	A form is submitted (Form submission) with the fields you define.
Schedule Trigger	A recurring schedule fires (Scheduled event) — minutes, hourly, daily, weekly, or monthly.
Webhook Trigger	An external system calls the agent’s webhook URL.
Email Trigger	An inbound email arrives at the agent’s address.

In the library, trigger filters label these as Chat message, Form submission, Scheduled event, Webhook, Email, and Manual. A Chat Trigger always makes the agent a Chat agent; Form and Schedule triggers are typical of Automation (pipeline) agents.

AI Steps

AI steps are where the model reasons, responds, loops, and branches. Add them from the AI Steps group.

Step	What it does
AI Respond	Generates a reply to the user. The default step for an Assistant after its Chat Trigger.
AI Think	Reasons over the input and produces an intermediate result for later steps, without replying to the user.
AI Agent	Runs another saved agent as a step, then returns its result. Used by Workforce agents to delegate to teammates.
AI Plan & Execute	Breaks a goal into a plan and works through it, calling tools and skills as needed.
AI Wait for User Input	Pauses the run and asks the person for more information before continuing.
AI Branch	Picks one of several paths based on the model’s read of the situation.
AI While	Repeats a set of steps while a condition the model evaluates stays true.
AI For Each	Runs the same steps once for each item in a list (for example, each overdue invoice).
AI Code	Runs a script over the step’s input and returns the result. Either the exact script you saved, or one the AI writes each run from your instructions.
AI Approval	Pauses for a person to Approve or Reject before the run continues. This is the human-in-the-loop step.

AI Code

Use AI Code for a deterministic transformation the model should not improvise: reshaping a payload, totalling lines, deriving a field. It takes one JSON object in and returns JSON out. It has two modes, set on the Code section of the step:

Run Code runs the exact script you saved. No model is called, so the step costs nothing and behaves identically every run. Reach for this by default.
Write & Run has the AI write a fresh script each run from your instructions, then runs it. Use it when the work varies per run, or when the step sits inside AI Plan & Execute and the plan decides what the code should do. It needs instructions and a model. A saved script is optional here — it becomes a starting point the AI adapts. The script it actually wrote is kept with the run.

Your script defines main(inputs, ctx) and returns JSON-serializable data. Choose JavaScript or Python on the step:

export async function main(inputs, ctx) {
  const overdue = inputs.lines.filter((line) => line.daysLate > 0);
  return { count: overdue.length, total: overdue.reduce((sum, line) => sum + line.amount, 0) };
}

def main(inputs, ctx):
    overdue = [line for line in inputs["lines"] if line["daysLate"] > 0]
    return {"count": len(overdue), "total": sum(line["amount"] for line in overdue)}

What the step receives

inputs is the JSON you mapped on the step. Leave the mapping empty and the step receives the previous step’s output instead — which is what you usually want in a straight-line agent. A mapped value that is exactly {{ … }} keeps its type, so a list stays a list rather than becoming text. A mapped reference that does not resolve stops the step. ctx holds outputDir, the folder a script writes files to when it needs to return more than JSON — a rebuilt spreadsheet, say. Files written there are saved with the run and passed on to later steps.

Limits

What	Limit
Each attempt	60 seconds
Write & Run overall	5 minutes, across up to 3 repair attempts
Input and result	4 MB each
Files a script can write	20 files, 10 MB each, 20 MB total

In Write & Run, a candidate that fails is shown its own error and given another attempt, up to the repair limit. Run Code never retries — a failing script fails the step.

The sandbox has no connectors, credentials, internet access, or memory between runs. Read from NetSuite in an Action step, pass the result into AI Code, and write back in another Action step.

AI Approval is how you put a human in the loop before consequential work. Use it ahead of any step that writes to NetSuite, and keep it in place rather than relying on Auto-approve writes. See Approvals.

Actions (By Connector)

The By Connector group lists actions exposed by your connectors — read and write operations against external systems. The exact actions you see depend on which connectors are connected to your workspace.

Source	Examples of what an action can do
Erstan	Load files, analyze datasets, create or update documents and files, send Erstan email, work with tasks and projects, and run code with Run Code or Write & Run Code.
NetSuite	Read records and run SuiteQL; create or update records such as invoices, vendor bills, and sales orders.
Other connectors	Post a Slack message, call an MCP server tool, or use provider tools such as web search.

See Tools overview, Tools by Erstan, and Tools by connector for the tool catalog.

NetSuite actions that create or update records are writes. Each write tool has a write policy of Allow, Require approval, or Deny, and runs use per-user authorization. Gate writes behind Require approval or an AI Approval step. See NetSuite and Tools, skills & connections.

An action step that runs its tool directly can carry an approval description (actionApprovalDescription): text shown on the approval card above the buttons, so the approver reads what the step is about to do. It may reference the output of earlier steps only ({{nodes.<id>.output.field}}), not the start step, the form or the trigger, and a reference that does not resolve stops the step before any card appears.

Logic and Utility

The Utility group holds structural and logic steps that shape the flow without calling the model directly.

Step	What it does
Note	A comment on the canvas to document the agent for builders. Does nothing at run time.
End	Marks where a run finishes. Pipeline (Automation) agents typically end here.
Rule-Based Decision	Routes the flow down a path using fixed, deterministic rules you define (not the model).
HTTP	Calls an external HTTP endpoint and returns the response.
Transform	Reshapes data between steps — map, rename, or format fields.
Extract	Pulls structured fields out of text or a document (for example, fields from an invoice).
Guardrail	Checks output against a rule and blocks or flags it when the rule is violated.
Skill	Loads a skill so the agent follows its packaged instructions for a task.

Use Rule-Based Decision when the choice is clear-cut and should always behave the same way, and AI Branch when the path depends on judgment about messy input.

Next steps

The visual builder

Place these steps on the canvas and connect them.

Triggers

Set up how each agent starts, including Run As.

Tools, skills & connections

Give actions the connectors and skills they need.

Choosing AI steps

Compare AI Think, AI Respond, AI Plan & Execute, AI Agent, approvals, branches, and loops.

Approvals

How human-in-the-loop checks gate NetSuite writes.

Get started

Chat

Agents

Apps & data

Skills & prompts

Solution Library

Tasks & teams

Artifacts & content

Playbooks

Plans & billing

Administration

Security & trust

Troubleshooting

Step reference

Triggers

AI Steps

AI Code

What the step receives

Limits

Actions (By Connector)

Logic and Utility

Next steps

The visual builder

Triggers

Tools, skills & connections

Choosing AI steps

Approvals

​Triggers

​AI Steps

​AI Code

​What the step receives

​Limits

​Actions (By Connector)

​Logic and Utility

​Next steps

The visual builder

Triggers

Tools, skills & connections

Choosing AI steps

Approvals

Triggers

AI Steps

AI Code

What the step receives

Limits

Actions (By Connector)

Logic and Utility

Next steps