The builders' agents
Each builder in Chartnaut has an agent in its chat, working on the one thing open in front of you. The indicator, definition, study and app builders all put the chat on the left and the code and preview on the right, so you watch each change land.
Use Agent in the sidebar for work that spans several pieces, such as an idea taken from definition to study to chart. Use a builder's chat to shape one piece until it is right. How the agent works covers the main agent.
Talking to a builder's agent
On a new piece the chat box asks you to describe it, and once there is code it asks What should change? Describe the change with specifics: a date, a bar, a number you expected.
Type @ in the chat box to hand the agent context from your library, such as another indicator, a definition, a playbook or a trade.
The model picker in the chat box sets which model does the work, and each message runs on AI credits or your own AI subscription, the same as the main agent.
How each builder handles changes
Builder | Code changes | Asks you before |
|---|---|---|
Indicator | Applied straight to the code | Using an indicator from your library |
Definition | Applied straight to the code | Using an indicator from your library |
Study | Proposed as a Study edit proposal | Every edit, and every run |
App | Applied straight to the code | Nothing: it publishes once its checks pass |
In the indicator and definition builders, each edit shows as a line in the chat listing the files it changed. When the agent wants to use an indicator from your library, it asks with an Indicator proposal card: Add VWAP?, with the settings it chose. Approve adds it; Reject opens a box for what should change.
In the study builder, each edit waits for you. The Study edit proposal card shows the change: Preview runs it before you decide, Accept saves it, and Reject sends it back with an optional reason. A run arrives as a Study run proposal, where you pick the run window and click Run.
In the app builder, the agent writes the code, checks the app on real data and publishes it as the last step of a build. Preview as shows a view on a different trade, day or report while you talk it through.
Asking the agent about the preview
When the preview hits a problem, a banner above it names it, often with a button that hands it to the agent:
Fix with agent on a runtime error, a script too heavy to run, or a preview with no visible output.
Diagnose with agent when the preview is waiting on data, or an indicator the script depends on returns nothing.
Ask agent on other warnings in the definition builder.
The button sends the banner's details to the chat, so the agent starts from the error itself instead of your description of it.
When you open a piece that belongs to someone else with Try it out, the banners show without these buttons.