Before you start
An agent belongs to a project, and anyone in the project except a viewer can create one. Your plan sets how many agents a workspace can hold. See Members and roles and Credits and billing.
Open Agents in the sidebar, or go to https://app.orbitalaip.com/agents.
Create an agent
Click New agent
Orbital creates the agent at once, names it New Agent and opens its conversation. If the project already holds an unfinished setup, New agent opens that draft instead.
Say what its job is
The agent asks what its job should be, with a few suggested replies and room for your own (Type your own answer…). Say what it should do, with which data, and what it should report.
Answer the second question, if it comes
When the job builds or tests strategies, the interview asks whether the agent should check with you first. Any other answer puts a permission on the card. See Give a permission from Chat.
Apply the card
The agent proposes its name, type, model and instructions on one card, titled Rename agent and the new name. Review it, then click Apply, or Discard to drop the proposal. A schedule is not part of the card: add one under Automations once the card is applied.
Until the card is applied, the agent cannot run. Anything that tries, such as Run under Test input, is refused with Finish and apply this agent's guided setup before running it.
Setup configures the agent; it does not do the job, which the agent does in its own runs.
Answers that give the interview enough to work with
Open Agent settings
Open the agent's menu
In the Agents list, hover the agent and click ⋮ (Actions for its name).
Choose Settings
The panel opens beside the conversation under the heading Agent settings. On a window narrower than a tablet it does not show; widen the window first.
Changes save as you make them, and ✓ Saved confirms each one.
Name, model and instructions
Name renames the agent.
Model is the agent's own, and a draft starts on GPT-5.6 Luna; when the agent has tools, the list shows only models that support tools. See AI models.
Instructions tell the agent what it does, how it analyzes, what matters and what is noise. Leave out rules about approvals and limits: what the agent may do on its own is set by its permissions, not its instructions.
Agent type, permissions, automations and knowledge
Agent type shows the kind in words. A new agent is a Job: Runs when you ask, or on a schedule. Manage opens Advanced settings, where Kind also offers Mission and Monitor.
Permissions says what the agent may do on its own, and only workspace owners and admins can change it. See The four permissions.
Automations → Manage opens the schedules and wakes that run the agent when nobody is here; schedules created there start paused. See Automations.
Knowledge → Manage binds Library documents and artifacts the agent can read and cite, each with its own switch. See Knowledge.
Advanced settings
Open the Advanced settings line, which summarizes what sits behind it, for example Focused · all tools · no output contract.
Creativity
Precise, Focused or Creative. New agents are Focused.
Reply length
The most the agent may write in one reply, in tokens: 32,000 for a new agent, or its model's maximum where that is lower. You can set any number up to the model's maximum. An agent with no tools and JSON output is capped at 16,000 tokens outside a conversation.
Memory and limits
Remembers names the memory the agent keeps across runs; every run reads what the last one wrote. Leave it blank for no memory.
Spend cap is how many credits one run may spend. A run that reaches it stops before its next step and keeps its work. Outside a task, calls on your own provider keys do not count toward it. Blank, shown as no cap, means the workspace limit is the only ceiling.
Step limit is how many times the agent may use a tool before it has to answer. Blank means 100, and the most you can set is 1,000. Reaching the limit is not a failure: the agent answers from what it gathered and says what it could not finish. The exception is a run started by a workflow's Agent node, which fails at the limit. The limit only applies to agents with tools.
Tools
Use all available tools is on for a new agent: every tool, including ones added later. Turn it off and the tools it covered stay ticked, grouped by category; untick the ones the agent should not use.
Hosted tools are Web search, on by default, and Code Interpreter, off by default. Once Make trading decisions is set, in any mode, the agent's runs get neither, whatever these boxes show. See Make trading decisions.
Output
JSON or TEXT; new agents answer in JSON. An Output contract, built with Add output schema or Edit as JSON, makes the answer's shape enforceable. Required top-level keys, such as verdict, confidence, are enforced even without a schema. With TEXT and a contract set, the final answer still has to be JSON.
Test input
Test input runs the agent on a sample payload, such as { "symbol": "AAPL" }. Run starts it in the agent's conversation. Save keeps the payload under a name, up to 12 per agent.
Lifecycle
Archive stops the agent and keeps everything; it then sits under the Archived filter. Unarchive brings it back as a draft, with its schedules off until you turn them on again under Automations.
Delete moves the agent to Trash, where you can restore it until it is permanently deleted after 30 days. See Trash.
Only project editors and above, and workspace owners and admins, can archive, unarchive, delete or restore an agent.