Skip to main content
An agent is only as capable as what you give it. Skills teach it how to do a job the expert way, tools and connectors let it act in NetSuite and other systems, and write approvals keep you in control of every change.
The visual agent builder showing a Chat Trigger connected to an Erstan action node, with the Add Step panel

The visual agent builder

​
Skills vs. tools vs. connectors

These three work together but do different things. Set them on an agent’s steps in the visual builder.
Think of a skill as judgment (when and how to do something) and a tool as a capability (the actual action). A good agent usually needs both — a skill that knows your AR collections process, plus NetSuite tools to read invoices and draft outreach.

​
Adding skills to an agent

Skills give an agent repeatable know-how — your dunning policy, how you categorize vendor bills, the exact SuiteQL pattern your team trusts.
1

Open the agent in the builder

From the Agents library, open the agent you want to edit. Select the AI step (for example AI Respond or AI Agent) that should use the skill.
2

Attach the skill

In the inspector, add the skill from your workspace library. Only Published skills (marked Agent ready) appear — a Draft skill is not usable by an agent yet.
3

Save and preview

Save the agent, then Preview a run to confirm the agent follows the skill’s guidance before you publish.
If the skill you want is still a Draft, publish it first. See Publishing, sharing & lifecycle. To build a new one, see Creating and editing skills.

​
Selecting connector tools

Tools are the actions an agent can actually take. They come from your connectors:
  • NetSuite — read records, run read-only SuiteQL, and (with approval) create or update records such as invoices, vendor bills, and customer messages.
  • Built-in Erstan tools — utility actions that need no setup.
  • Other connectors — Slack, MCP servers, and more (see Connectors overview).
In the builder, connector actions live under the By Connector node category. Add an action step, then pick the connector and the specific tool it should run. For the full catalog of step types, see the Step reference. New Erstan actions use Run action: configure the inputs and the operation executes directly. Use AI Plan & Execute when a model should choose inputs or combine tools. File extraction and Write & Run Code may use models inside their operation; ordinary parsing and fixed code do not need an outer planning step. For the source-based tool catalog, see Tools overview, Tools by Erstan, and Tools by connector.

​
Run Code and Write & Run Code

Two built-in Erstan actions run a script as a tool, so a tool-using step can compute without you adding a separate AI Code step. They behave like the AI Code step — same two modes, same sandbox, same limits — and you configure the script on the action itself. The difference is where the input comes from, set on the action’s Input control:
  • Agent — the model composes the JSON each time it calls the action. Flexible, and the right default when the input depends on what the step just worked out.
  • Manual — you map the input yourself. Erstan fills it in from the run before the action starts, so the model never has to write the data out.
Use Manual when the input is large. A few thousand rows mapped manually go straight to the sandbox, while the same rows in Agent mode have to pass through the model first — slower, costlier, and bounded by how much the model can reliably reproduce.

​
The “Connection required” banner

A step can only run if its connector is connected and authorized. If something is missing, the agent shows an amber Connection required banner, and the run is blocked until you resolve it.
An owner or admin sets up the shared workspace connection once. Open Connectors and connect the system the step needs.
NetSuite uses per-user authorization — each member authorizes their own NetSuite user and role before agents can act as them. Authorize from the connector’s Setup tab, or from the Authorization Required card in chat. See Authorizing your NetSuite user.
Scheduled and form-triggered agents run under a specific user. If that identity isn’t configured or consented, the builder flags Trigger setup and publishing is blocked. See Preview, publish & versions.

​
Write-capable tools need approval

Reading data is safe. Writing to NetSuite — creating or updating records — changes your books, so Erstan gates it. Every NetSuite tool has a write policy an owner or admin sets on the connector’s Tools tab:
Write-capable NetSuite tools require approval by default. When an agent reaches one, it pauses and shows an approval card with a before/after preview — nothing is written until you approve. Loosen this only when you fully trust the action. Manage write policy per tool on the Managing NetSuite: tools & write policy page, and review how approvals work in Approving and rejecting AI actions.
You can add an explicit AI Approval step in the builder to force a human check at a specific point in a run, even for actions that would otherwise be allowed. See the Step reference.

​
Why Code actions don’t ask for approval

Run Code and Write & Run Code are classified read-only and run without an approval pause, even though Write & Run Code executes a script the AI just wrote. That is deliberate: the sandbox has no connectors, credentials, internet access, or way to change anything outside the run, so the script can only compute over the input it was handed. The write is the step that comes after. Put the AI Approval step in front of the Action that saves the result, not in front of the code that produced it.

​
Next steps

The visual builder

Add steps, attach skills and tools, and resolve connection warnings.

Step reference

Every step type, including By Connector actions and AI Approval.

Managing NetSuite tools

Set write policy (Allow / Require approval / Deny) per tool.

Security & write safety

How per-user authorization and approvals keep your data safe.