Skip to content
RADIAN/ DOCS

Modes and autonomy

Choose how an agent should work and control which actions can proceed without interruption.

Radian separates the shape of the work from the approval policy applied to it.

Modes

Build is the default and stays visually unmarked. Other modes change the capability envelope:

  • Plan develops an implementation approach before changing code.
  • Ask focuses on explanation and investigation.
  • Advisor provides judgment without taking over execution.
  • Goal sustains work toward a declared outcome.
  • Multitask coordinates multiple lines of work.
  • Debug emphasizes diagnosis and evidence gathering.

The interface teaches these modes through the composer rather than adding a permanent row of controls.

Autonomy

Autonomy is governed by a safety floor, remembered allowances, and the selected policy. An approval can remain a one-time decision or become an explicit “always allow” rule. Actions prohibited by the safety floor remain blocked unless the corresponding product setting explicitly permits them.

Direct shell commands entered by the user are treated as human actions, not agent approvals.

Interruption

Escape follows a deliberate stack: close the current overlay, close transient shell UI, interrupt the running agent, then clear the composer on a second action. This keeps one key predictable without assigning it several simultaneous effects.

On this page