The Autonomy Dial: From Read-Only to Full Control

The question that matters with agents is not "how autonomous can it be" but "can I set exactly how autonomous it is, per task". In idapt that setting is the autonomy dial: four levels, per chat, enforced by the platform rather than promised by the model.

The four levels

• Read-only: mutating tools are stripped from the model entirely. The agent can look (read files, search, browse) and point, but nothing it does changes state. The mode for exploration and audits.
• Ask: the agent works, but consequential actions pause for your confirmation inline: Allow once, or deny. The mode for new agents and unfamiliar tasks.
• Auto (the default): routine actions proceed; boundary-crossing and irreversible ones still confirm. The everyday balance.
• Full: the agent runs to completion within its budgets. The mode for trusted, repeated workflows and scheduled automations.

Defaults live per agent; workspaces can set a ceiling no chat may exceed.

Enforcement, not etiquette

The dial is not a system-prompt suggestion. Read-only removes the tools from the model's palette; confirmations gate execution out-of-band; budgets cap spend server-side. A persuasive prompt cannot talk the platform out of the level you set. (It is an oversight mechanism, not a security boundary against a malicious human: those are permissions, which are separate.)

Around every level, three constants:

• Budgets: a run stops when its spend cap is reached.
• Traces: every action a run took is inspectable afterward.
• Per-computer consent: screen control requires explicitly enabling that computer, whatever the dial says.

Choosing a level, practically

New agent or new task type: Ask, until its judgment has a track record. Daily driving: Auto. Scheduled work you have watched succeed: Full with a tight budget. Auditing something delicate: Read-only, where the worst case is a wasted question.

Good to know

• Confirmations arrive inline at the action, with a compact review banner when several are pending.
• Agents can escalate their own actions to confirmation when instructions say to; they cannot de-escalate yours.
• The help article documents every rule the dial applies.

Set your next unfamiliar task to Ask and watch what the agent requests: ten minutes of confirmations teaches you exactly where your comfort line is. The agents overview covers the permission model underneath.