What an agent is
An agent is a saved worker in your project. It has Instructions, a Model, a set of tools, permissions, and a conversation of its own.
You talk to it in that conversation. Each reply is a run: the agent reads your message, calls tools until it has an answer, and writes the answer back.
A new agent has Use all available tools on, which gives it the workflow nodes as tools, market data included. Web search is on, and Code Interpreter is off until you turn it on. Switch tools off in its settings to hold it back.
The Agents page
Open Agents in the sidebar. The left pane lists your agents and channels. Open conversations show as tabs on the right; drag a row from the list onto them to see up to four at once.
New agent creates an agent at once and starts a setup interview in its conversation. See Set up an agent.
The list has two filters, All and Archived, and groups its rows:
- Channels: shared conversations, with their member avatars.
- Needs you: agents waiting for your answer. The badge counts the approvals waiting.
- Working: agents with a run in progress or a task under way.
- Watching: agents waiting on a signal.
- Recent: everything else, latest activity first.
Hover a row and click ⋮ for Open, Settings, Appearance, Archive and Delete. A channel row's menu has Open, Archive and Delete.
The conversation
Type in the composer and press Send message. Stop response (■) stops the reply while the agent is answering, and Show work under a reply opens the run behind it.
A run can also start from a schedule, a signal the agent watches, a workflow's Agent node, another agent, a task or a channel. Every run lands in the agent's conversation. One that is not a reply to you appears as a card with a status such as Finished, Working, Stopped or Didn’t finish. Expand it to read INPUT, PLANNING, WORK (each tool call), EVIDENCE (the sources it cited) and OUTPUT. A card that is still working has Stop.
When a run files a document, the OUTPUT section shows it with Open and Add to Library. Its More (⋯) menu holds Download Markdown, Download JSON, Download CSV for tables, and Add to a board for a dashboard.
Beside the composer you can Attach images and documents, open Automations for the agent, or Give the agent a task. A task is a goal it keeps working at across runs until it says it is done, gets stuck or uses its budget. See Tasks and Automations.
Notifications tell you when a run you started fails, when an unattended agent waits on your approval, and when a task finishes or needs you. See Notifications.
What it can do on its own
Within its tools, an agent reads data, searches the web, changes watchlists and sends alerts to the person who created it. It creates workflows, edits the drafts it made, builds programs, and reads strategies, backtests and deployments. It can also create draft agents, add schedules that stay off until you turn them on in Automations, and hand a task to another agent.
Every run is bounded by the agent's Step limit, 100 unless set, and its Spend cap in credits. See Memory and limits.
Each run's backtests, program builds, new workflows, workflow edits and deployment changes are capped, and so are the backtests an agent starts on its own each day. See What an agent can do.
With Make trading decisions set, in any mode, an agent keeps only read tools such as prices, news and watchlists, and a task it is working on pauses. See Make trading decisions.
What it must ask for
Build and test strategies, Make main, Make trading decisions and Supervise deployments decide what an agent may spend or change unasked, and only workspace owners and admins set them. See The four permissions.
Under Ask, the act waits in the conversation as a Waiting for you card for up to seven days, and the agent moves to Needs you. Approving a deployment change needs a workspace owner or admin. See answering a request.
A trade under Ask waits for review in Trade instead. See Approvals and permissions.
No agent action deploys a strategy to Trade or retires one; a person does that.
Agents, Chat, Workflows and strategies
Chat is a conversation with Orbital, and it can create, edit, run and schedule agents for you. It can also propose Build and test strategies for a workspace owner or admin to apply; the other permissions are set in Agent settings. See Chat and Give a permission from Chat.
Workflows run a graph; agents run a loop. A workflow calls an agent through its Agent node, and an agent can create workflows. See Workflows.
An agent can write a strategy as rule cards or a program, run backtests and read the results. It can copy a deployed version into a new draft, and make a version main when Make main allows it. See Agents and strategies.
Channels
The # button (Create new channel) creates a channel of up to eight agents. A plain message reaches every member, and a lead, when the channel has one, sums up their answers. See Create a channel.