Overview
Chartnaut's agents write the same scripts this section documents, and check them with lint, typecheck and a dry run before you see the result. This page shows which agent writes what, and how its changes reach you.
The product side, with every card and button, is in How the agent works, Approvals and The builders' agents.
Who does what
Agent | Where you find it | Writes | How a change reaches you |
|---|---|---|---|
The Agent | Agent at the top of the sidebar | Indicators, definitions, studies and apps, often several from one request | Edits land in private drafts straight away; it stops at approval cards |
The chat in the indicator builder | One indicator | Applied straight to the code | |
The chat in the definition builder | One definition, every file of it | Applied straight to the code | |
The chat in the study builder | One study | A Study edit proposal you accept or reject, and a run card you approve | |
App builder's agent | The chat in the app builder | One app | Applied straight to the code, and published once its checks pass |
Inside the Attach to chart wizard | The | Runs when you click Apply breakdown |
The Agent and the builders
The Agent does the whole build itself. It is taught the same rules as the indicator, definition and study builders' agents: their DSL references, patterns and pre-apply checklists are included in its skills word for word. It checks its work on the server (lint, derive, dry run, then a real run) instead of in a builder's chart preview.
What it builds is an ordinary indicator, definition, study or app. Open it in its builder and carry on there with that builder's agent. A study the Agent built opens in the study builder with a note in its chat saying where it came from, so you can keep refining it.
Use the Agent for work that spans several pieces, such as a definition, its run over history, and a study on its events. Use a builder's agent to shape one piece until it is right.
Supervised and Autonomous
The mode capsule in the Agent's chat box sets when it shows you its work.
Supervised shows each piece on a Result card as soon as it checks clean, and always before anything uses it: a definition before it runs over history, a study before its run.
Autonomous shows the finished piece once, at the end.
In both modes it stops for a Plan proposal, a Data collection proposal, a Study run proposal, a Profile override proposal and any question only you can answer. An approved run that fails on a code bug is fixed and run again without a new card; a new card comes only when the scope changes, such as a different instrument or a wider range.
Rules every agent follows
Every exact string is looked up, never remembered: slugs, output ids, accessor keys, param ids and version numbers come from a tool result. A plausible but wrong string passes lint and fails silently on real data.
Every dependency is pinned to the version a search or approval returned. A clean edit can create a new version, so the version read before an edit is already out of date.
Checks are climbed in order, and each proves only its own claim: lint, then typecheck or derive, then a dry run, then a real run. See Agent lint and dry run.
No agent can see the rendered chart. You are that check, and every result says so under Not checked.
Agents never publish or share an indicator, definition or study. That is your button. The one exception is an app, which the app builder's agent and the Agent publish once its checks pass, because an unpublished app shows nothing.
The study builder's agent does not write Forward Insight code. The Forward Insight agent does, inside the attach wizard.