What an agent can do
An agent works on strategies in its own runs, on the same records Backtesting and Trade show you.
| Act | What it does | Needs |
|---|---|---|
| Read strategies, runs and deployments | Lists strategies, reads one strategy's versions and runs or one backtest, and lists and reads deployments. | Backtesting active in the project for strategies, Trade for deployments |
| Create a Rules strategy | Saves a workflow as a draft. | Nothing |
| Edit its own draft | Replaces the whole graph. | Nothing |
| Branch a deployed strategy | Copies the version a deployment runs into a new draft. | Nothing |
| Build a program | Compiles a Program node's Rust source into the module a backtest runs. | Nothing |
| Run or cancel a backtest | Prices the run, then asks you or starts it. | Build and test strategies |
| Make a version main | Moves the strategy's main mark to the version it names. | Make main |
| Control a deployment | Pauses, stops, resumes or re-limits a deployment. | Supervise deployments |
No agent act deploys a strategy or starts a forward test; you do those yourself in Backtesting and Trade.
One run may create two workflows, branches included, and make six edits and six builds. It may ask for three backtests and six deployment controls, and ask once to make a version main.
An agent with Make trading decisions set, in any mode, is given none of these tools. See Make trading decisions.
Permissions and approval requests
Backtests need Build and test strategies, moving main needs Make main, and deployment controls need Supervise deployments; only workspace owners and admins set them. See The four permissions.
Under Ask, the act waits as a card with Approve and Refuse for seven days. See answering a request.
Reading strategies, runs and deployments
An agent's strategy list gives the 50 most recently updated strategies, and one strategy read adds its newest 25 versions and newest ten runs. The agent is told to quote a backtest's Sharpe ratio only with the number of configurations tried and the deflated figure, and to state only numbers it has read. A deployment read carries no profit and loss; Trade's Deployments list shows REALIZED GAIN.
Writing rule cards
The agent writes a Rules strategy's rule cards and their connections. Orbital lays out the graph, checks every required setting and saves it as a draft that runs nothing until someone runs or schedules it. Under the reply it shows as Workflow · draft, with Open.
When the graph qualifies, it appears at once in Backtesting's Strategies list as Draft, and turns Tested when a backtest completes. The agent may declare parameters a backtest can set without editing the graph.
It may edit only a draft it created, and only while nothing runs it; a published or scheduled workflow, or one a person made, is refused.
Branching a deployed strategy
While a deployment runs a strategy, the agent may not change it. It branches instead: a new draft holding the version the deployment runs, named after the original with branch added. The deployment keeps running its own version. Changes saved after the deployed version are not in the branch.
The branch becomes a strategy on its own when its graph qualifies, and the agent edits and backtests that. It shows as Workflow · branch.
Building a program
The agent saves the Program node with History the program reads (bars of the decision bar) set, writes the Rust source and builds it. Failed builds count toward its six a run. See A Program from Claude or an agent.
Running backtests
The checks run in this order, and nothing is spent before the last:
- The permission. Deny refuses before anything is compiled.
- The estimate. The strategy is compiled and priced with the Run dialog's data check. A compile or data refusal comes back as an answer to fix, never as a card.
- The spend cap. A run the agent's Spend cap cannot cover is refused before any card or run exists. Backtests this run started on its own count against the cap; one you approve is your spend.
- The card or the start. Under Ask the run waits as a card titled Run a backtest of the strategy's name, whose detail gives the symbols, bar size, window and cost. Under Always approve it starts, and the credits are attributed to the agent.
The agent picks the window as a lookback such as 3y or 90d, or as dates; the end defaults to yesterday. Capital defaults to $100,000 and margin to off, and the cost profile is Alpaca's unless it picks zero. It may also set the declared parameters and a one-line note on what the run changes; the version its backtest makes carries that note.
A configuration that already ran over the same window is refused with that run's id, so nothing is spent twice.
A run may ask for three backtests, counting cards still waiting for you. On its own, an agent may start 24 a day, resetting at midnight US Eastern; backtests you approve do not count. In your own conversation, or on a task, the agent is woken when a backtest finishes and reports the result. A scheduled run is not woken; the result waits in Backtesting.
Making a version main
Under Ask the card reads Make v2 main for the version the agent names, and a run may ask once. It reads No longer applies if a Rules or Program version has no completed backtest or main moved after it asked. See Main.
Controlling a deployment
The agent can change Budget, Max per trade, Open positions and Daily loss pause; a limit it does not name keeps its value. Each card is titled with the act and the strategy's name.
| Card | What approving does |
|---|---|
| Pause | Nothing new is placed until it is resumed. |
| Stop | It does not trade again. |
| Resume | It then trades on its own as before. |
| Re-limit | It then trades on its own within the new limits. |
The broker account's own setting, Ask, is the ceiling on the agent's permission, so under Ask, Auto or Always approve every act waits as a card, and Deny refuses. See Supervise deployments and Manage a deployment.