Agent preferences

The settings panel on the right of the agent builder has three tabs: Preferences, Advanced and History. This page covers the first two. Memory and chat history are on their own page, tools on another, and History is in share and publish.

Each section is a collapsible band with its own icon and a one-line summary of its current value. Changes save as a draft about a second after you stop typing. Use the undo and redo buttons in the top bar, or Ctrl+Z and Ctrl+Shift+Z (⌘ on a Mac), to step through your edits. Undo history lasts while the builder page is open.

Preferences tab#

Details#

This section is folded when you open the builder. It holds the facts about the agent rather than how it answers.

• Name. Up to 100 characters. The name saves when you press Enter or leave the field.
• Description. Shown on the Agents list, the library card and the chat welcome.
• Category. The category the agent is filed under.
• Chat input placeholder. The grey text in the empty chat box.
• Disable text input. Turns the chat text box off.
• Detailed description. A longer text, written in Markdown.

Model & instructions#

• Model. The language model the agent thinks with. The default selects the model your workspace marked as default.
• Reasoning Effort. Appears only for models that support it. The values on offer depend on the model. The default is none.
• Deep agent. A switch. It gives the agent a file system, a planning tool and one helper sub-agent it can hand work to. Use it for long, multi-step tasks; a plain agent answers faster.
• System Message. The instructions. Say who the agent is, what it may do and how it should answer. The default is “You are a helpful assistant.”
• Show Progress. Shows the agent’s progress and tool calls in the chat while it works. It is on by default.

If you switch to a model that does not accept the chosen reasoning effort, the builder resets it to none, or to the model’s first allowed value.

Apps#

The tools the agent can call. See agent tools.

Triggers#

Chat is always on: it covers the test chat, chatbots and the API. Click Add trigger to let something else start the agent, such as a Slack message, a schedule or a webhook. Each trigger has a status badge:

• Active: the trigger is live.
• Publish to activate: the draft has a trigger that the published version does not. Webhooks and schedules register when you publish.
• Connect: the trigger needs an account you have not connected. Click it to connect. See connecting accounts.

The gear opens the trigger’s settings. The three-dot menu removes it. The available triggers are listed in the step reference.

Advanced tab#

Generation#

• Temperature. Higher values make answers more random, lower values more deterministic. The default is 0.7 and the range is 0 to 2, but some models cap it lower (Anthropic models stop at 1). It is hidden for reasoning models, which ignore it.
• Max Output Tokens. The longest single answer. The default is 3000.
• Max Iterations. How many rounds of thinking and tool calls the agent may run. The default is 20 and the maximum is 250.

Structured output#

Switch it on to make the agent answer as JSON in a shape you define. In Output Schema you list each output field with a name, a type (string, int, float, bool, list or dict) and a description.

When chat opens#

Blocks the chat shows before the first question, such as a welcome message, buttons or cards. Click Add block and pick one from the list. Drag blocks to reorder them, and remove one from its row. Open a block to edit its content. With no blocks, the chat opens empty.

Chat history#

See memory and chat history.

If something goes wrong#

If you cannot edit anything, you may have opened the agent in View as a user mode or your role may lack edit rights. Click the pencil Edit agent button in the top bar. If the panel shows that the flow’s nodes no longer match the agent form, see convert to workflow.