Building the automation with AgentMaker

AgentMaker turns a plain-language description into a proposed flow. It reads the automation’s graph, the step catalog, recent runs, and organization references, builds the whole graph, and validates it before showing the proposal.

It lives in the AgentMaker dock, on the right edge of the screen, and follows you across screens. The outcome depends on where you are:

with that automation’s editor open, the proposal becomes a Draft review: applying it changes only the canvas, and saving is still your decision;
on any other screen, the proposal becomes a Proposal for you to confirm, and Confirm saves a new version that goes live immediately.

Where to open it and who can use it

In the automation editor, select AgentMaker in the top bar (tooltip: “Build this automation by chatting with AgentMaker”). The button only appears while the dock is closed; with the dock open, use the dock itself. Outside the editor, open the dock from the AgentMaker tab on the right edge.

When open, the dock pushes the page aside instead of covering it, and it can be widened or narrowed. On a phone, it takes the full screen.

Using the AgentMaker requires permission to edit agents (agents.write), and changing an automation requires automations.write; without both, the editor button does not appear. That permission does not, on its own, allow creating tags or lists; see When a tag or list is missing.

Write a verifiable request

State the event, filters, timing, effect, and resources to use. For example:

When a conversation enters the Support inbox, wait 1 hour. If no one has replied, hand it off to the Triage agent.

If a decision that changes the outcome is missing — the inbox, deadline, message, or branch criterion — AgentMaker should ask one question at a time rather than invent it. Use the names visible in SquadOS and say explicitly when you want to preserve part of the current flow.

Before showing any proposal, the backend runs the graph through the same validator used when saving. AgentMaker may correct an invalid answer and try again, up to three validation attempts in the same request, with at most 40 steps per flow. If it cannot, it should explain that it did not build the flow; a missing card is not an applied change.

With the editor open: Draft review

With the automation editor open, make the request in the dock. The Draft review card shows the added, removed, and changed steps and the connections created or removed.
Review everything and select Apply to draft. The graph moves to the canvas as an unsaved change. The state changes to “Applied to the draft — review it and click Save when it is right.”
If you do not like the result, Undo application restores the previous draft, as long as you have not changed it since. Discard abandons the review without applying it.
Review the canvas and every field marked as required or invalid. Validating the proposal does not replace that inspection; also confirm, in each step’s panel, resources such as inbox, agent, tag, and template.
Use Save only when the draft is correct. On this path, AgentMaker does not save, turn on, or turn off the automation.

If you edit the canvas after the review appears, the review does not overwrite your edit: the card shows “The draft changed. I kept both your edit and the proposal so you can review or recalculate.” Use Recalculate to rebuild the review on top of the current draft. A review replaced by a newer one can no longer be applied.

If the conversation is open in more than one tab, only one of them can apply or discard the review; the others show its state.

Outside the editor: Proposal for you to confirm

When you ask for a change to an automation without being in its editor, the Proposal for you to confirm card offers:

Confirm — saves the proposed graph as a new version. If the automation is on, that version takes effect immediately for upcoming runs;
Review on screen — opens the proposal in the editor so you can review and save it yourself;
Discard — abandons the proposal.

If the automation changed since the proposal was shown, Confirm is refused with “The proposal changed since it was shown. Ask AgentMaker to redo it.” Prefer Review on screen for active automations that talk to customers: the confirmed version goes live without passing through the canvas.

When a tag or list is missing

AgentMaker can create the contact tags and contact lists the flow needs. These creations happen right away, with no confirmation card, and it reports what it created.

A tag requires permission to edit contacts (contacts.write).
A list requires permission to edit emails (emails.write).

AgentMaker reads at most 300 tags, 300 lists, 300 inboxes, 300 agents, and 300 people per request, without name search. If an existing resource does not appear to it, do not ask it to create a duplicate: check the name in the corresponding catalog and set that reference manually in the step.

It does not create inboxes, channels, or agents. For those resources, it can only recommend the corresponding page; finish creating them outside the dock and then make a new request.

What is out of scope

AgentMaker for Automations cannot:

create a new automation — it changes the flow of an automation that already exists;
turn the automation on or off, or change its Settings;
run or test the flow;
create inboxes, channels, or agents;
guarantee that a real send will happen: channel windows, consent, frequency, opt-out, permissions, and data present at run time still apply.

It can read the automation’s recent runs to diagnose a problem, but the suggested fix follows the same review or confirmation path described above.

Conversation, stopping, and resetting

The dock conversation follows you across screens; the current screen is sent as context with each request. Reset chat asks for confirmation and starts over, without undoing what was already applied or saved, or removing tags and lists already created. The local copy of the conversation stays in this browser, per organization, on a best-effort basis: another browser or device does not restore it.

While a request is running, use Stop to halt it.
If you leave the area or resource during a request, it is interrupted and the dock says why.
After reloading the page, treat an old review or proposal as history if its card no longer offers actions. Ask for a new proposal.

Review and test

Before saving, trace the flow from the trigger to every exit, open each step, and confirm references, text, timing, variables, and connections. Then use Test. The test runs the current draft: sends are simulated, but AI and webhook calls are real. See Runs and tests before running it.