Agents

The assistant

The assistant is a chat panel on every Agents page. Open it with the Assistant button in the top bar or ⌘J (Ctrl+J). Describe what you want in plain words and it sets it up through the same Agents API the pages use:

• workflows, including the code for script steps;
• agents, skills, sandbox profiles, gateways, model policies, triggers and campaigns;
• starting, cancelling and explaining runs, and questions such as "why did my last run fail?";
• checking on your runners.

#Approving changes

The assistant reads freely, but every change waits for you. It shows a card with a one-line summary and, for resources, the plan (created, updated or unchanged, for each document) and the YAML. Choose Apply or Approve to go ahead, or Decline. If you send a new message instead, the waiting changes are declined.

Tick Make changes without asking to let it go ahead on its own. Either way it acts with your own permissions in this workspace, and everything it changes is in the audit log as done through the assistant.

It can't make API keys or registration tokens, change its own settings, or reach Docs, Projects and Chat directly: workflows do that. Never paste secret values into the chat; the assistant refers to secrets by name.

Your chats are kept with your account. Earlier chats lists them, and New chat starts over.

#Where it runs

The assistant's model calls run on one of your runners, never in SpudSuite's cloud. SpudSuite sends a runner one model call at a time, the runner calls the model and sends the reply back, and SpudSuite runs the assistant's tools as you. A runner can answer in two ways:

• Signed in with Claude. When you sign in with Claude in the runner app (the same sign-in Factory uses), the runner answers your own chats on your own Claude plan. It never answers anyone else's, because one person's plan is for that person.
• With an API key. A runner with an ANTHROPIC_API_KEY in its secrets answers anyone in the workspace. The key stays on the runner.

If no runner can answer you, the panel says what to do next and has a button to the page that does it: Add a runner when the workspace has none, the runner's page when it's offline, or its AI settings when it's online but Claude isn't set up on it yet. Your message waits until a runner can answer.

#Settings

Workspace admins choose the assistant's model, gateway and effort, and which runners answer it (by label), on the Assistant settings page. The gateway is either a provider name such as anthropic, or the name of a Gateway resource the runners know.