Radiant documentation
Agents

Permission modes

The three modes a chat can run in, what each allows, and why Auto is deliberately awkward to turn on.

Every chat runs in one of three modes, chosen from the control at the bottom of the message box. The mode decides whether the assistant can change your map, and whether it asks first.

The three modes

ModeThe assistant can readThe assistant can change your map
Read onlyYesNo, the tools that write are not offered at all
ManualYesOnly what you approve, one change at a time
AutoYesYes, without asking

Manual is the default, and it is the right setting for almost everything.

Read only

The write tools are not merely disabled: they are not given to the model, so it cannot attempt an edit and be refused. Useful when you want analysis of a map you would rather not risk, or when someone else is mid-edit.

Manual

The assistant can propose changes, and each proposal waits for you. You see what it wants to do in plain language before anything is written, and you can approve it, or reject it with a note about what to do differently.

This is the mode the approval flow is built for. Approving map changes covers what the review looks like. Question drafts need no second approval: asking for a question draft adds its node and starts the draft, and asking to change a draft changes it, as Drafting a forecasting question describes.

Auto

Auto removes the approval step. The assistant applies map and project changes as it decides on them. Starting a Metaculus AI forecast keeps asking regardless, even here; see Starting an AI forecast.

What Radiant tells you when you turn it on

Radiant will apply canvas and project changes without asking for each edit. Changes may be difficult to undo, and repeated execution can duplicate some imports. Only enable this when you trust the chat instructions and canvas content.

Two constraints make it less dangerous than it sounds:

  • One acknowledgement per chat. You confirm a warning dialog before it turns on.
  • This tab, this chat. The setting is stored per browser tab and does not follow you to another chat, another tab, or another device. Safe modes are remembered normally.

And one that does not change:

  • It is not extra authorization. Auto removes the asking, not the permission check. If you cannot edit a project, neither can the assistant acting for you.

The real cost is prompt-injection containment

Per-call approval is the last human check between untrusted content – a map imported from elsewhere, a web page the assistant read – and a write to your project. Auto removes that check by design. It is a reasonable trade on a map you built yourself, and a poor one on a map you did not.

Changing mode

The mode can only be changed while the chat is idle. Changing it applies from your next message; it does not retroactively affect a change already proposed or applied.