Skip to main content
Double-click an actor on the canvas to open its panel on the right. You can also select the actor and click Edit in its toolbar, or press Enter on it. A new actor opens by itself. This is the one place where you tell the actor who it is, what it knows, and what it should do. Routing is set on the routes themselves; see Routes.

​
Name and icon

At the top of the panel:
  • Icon — Click the icon to pick from a Lucide icon library. The platform also auto-suggests an icon based on the actor’s name and context. You can change it any time.
  • Name — Short and descriptive. Appears on the canvas and in conversation transcripts as the “speaker”. Good: Pricing Specialist, Returns Intake, Tier 2 Support. Bad: Agent A, Bot.
An actor needs a name. If you clear it, the panel says Give the actor a name. and stays open until you type one.

​
The sections

Under the name, a rail on the left of the panel lists the actor’s sections. Click one to show it, or use the arrow keys once the rail has focus. A number next to a section says how many items it holds. The panel opens on Instructions. When you switch to another actor while the panel is open, it stays on the same section. When the panel is narrow, the rail shows icons only; hover one for its name. Tool, knowledge base, integration, hook and step editors open inside the panel, in place of their list, with a back arrow to return to it.

​
Instructions

​
Context

Background knowledge the actor needs. Inheritable by children. See Context and inheritance for the full story. Up to 4000 characters; a counter under the field shows how many you’ve used.
Don’t repeat content that lives in Global Context. The actor inherits it automatically if Global context is checked under Inherit context.

​
Instructions

What this specific actor should do. Private — never inherited by children.
This is where you put the actor’s persona, its task, and any handoff rules. If the actor has steps, a note above the field says so: the actor runs its steps in order, and these instructions still guide every AI step. Click Steps in the note to go to them.

​
Definition of Done

A short rule for when the actor’s job is complete.
When the actor decides it has satisfied this rule, it signals completion. The platform then either hands off or closes the conversation.
Strict-mode actors must have a Definition of Done. Without one, the actor won’t be allowed to hand off — and the conversation will stall. See Focus modes.
If you leave it blank, the actor decides on its own when it’s finished. That’s usually fine for short, single-purpose actors. Set it explicitly for anything multi-step.

​
Focus mode

Auto / Flexible / Persistent / Strict. Controls how strictly the actor sticks to its job when users change topics. Default is Auto. See Focus modes for the deep dive.

​
Inherit context

Checkboxes that let this actor opt into upstream context:
  • Global context — the globalContext from Agent settings.
  • Slots collected so far — slot summary (default: on). See slots.
  • <Each ancestor actor> — any actor that routes into this one.

​
Writing in the prompt fields

Context, Instructions and Definition of Done are prompt fields. Type @ to mention a slot, parameter, tool, knowledge base or app, or click Insert @ in the field’s corner to pick one from a list. Expand opens the field in a large editor. A mention of something the agent no longer has is flagged under the field. See Prompting.

​
Steps

With no steps, the actor is free-form: it thinks and replies turn by turn, following its Instructions. Add steps in the Steps section to script the actor’s behaviour:
As soon as an actor has a step, it runs scripted: the steps top to bottom, and it’s done after the last one. Its Context and Instructions still guide every AI step. See Step builder for when to use each, and how to add, reorder and edit steps.

​
Tools, Knowledge, Integrations and Hooks

These sections give the actor its capabilities:
  • Tools — turn on web search, web extract or code execution, and add custom tools: run Python, call a workflow, hit an HTTP endpoint.
  • Knowledge — give the actor search access to your documents.
  • Integrations — let the actor send Slack messages, send Gmail emails, look up Shopify orders, etc.
  • Hooks — run a hidden action when something happens in the conversation, for example when a slot is captured.
See Capabilities for the full guide.

​
Saving

The actor’s own fields save automatically, a moment after you stop typing and whenever you leave a field. The header shows Saving…, then Saved. If a save fails, it shows Couldn’t save with Retry.
  • On Tools, Knowledge, Integrations and Hooks, each change saves straight away. An item’s editor saves it with its own Save button.
  • A step’s editor has its own Save too. See Edit a step.
When you’re done, close the panel with the X, or click an empty part of the canvas. Nothing waiting to save is lost. Your edits go to the agent’s draft. Tests run them straight away; customers see them once you deploy. A conversation that’s already running keeps the version it started with.
An agent that went live before versions existed shows a warning at the top of the panel and keeps a Save Changes button on Instructions, because its edits reach customers at once. The other sections still save each change straight away. See Agents that went live before versions.
If you’re iterating on a prompt, keep the Test drawer close by. Edit, start a new test, send the message again — you’ll see the new behaviour in seconds.

​
Delete an actor

Select the actor on the canvas, then either:
  • Click Delete in the toolbar above it.
  • Press Delete or Backspace on your keyboard, while no panel is open.
Deleting an actor also deletes its steps and every route connected to it. Nothing asks first: a toast offers Undo, and Ctrl+Z on the canvas puts it back too.

​
Anti-patterns

One actor that does five things. If your Instructions block is more than a couple of paragraphs, split into multiple actors with routes between them. Focused actors route better and produce better replies.
Instructions in Context. Behaviour rules (“always confirm before sending”) belong in Instructions. If you put them in Context, child actors will inherit and start trying to confirm too. Almost never what you want.

​
Next

Routes

Connect actors and define when handoffs happen.

Capabilities

Give the actor tools, knowledge, integrations and hooks.

Step builder

Script an actor’s behaviour step by step.

Agent settings

Configure flow-wide context, knowledge, and integrations.