Autonomy and confirmations
Found this helpful? Share it:
Found this helpful? Share it:
You decide how much an agent does on its own. The autonomy dial sets the default, confirmations let you approve or deny individual actions, and you can stop a run at any moment 🎚️
Every chat has an autonomy setting in the composer tools dropdown, the sliders icon next to the message box. It has four rungs:
Read-only: the agent can look at things but cannot change anything. Every tool that would edit a file, run a command, or touch a resource is removed from what it can reach.
Ask: the agent works on its own but pauses for your approval before anything consequential.
Auto: the default. The agent acts on its own and pauses only before something that would be hard to undo.
Full: the agent runs without pausing.
The dial is about oversight, not security. It changes how often the agent checks in with you. What an agent is allowed to reach at all is set by its permissions. See Agent permissions and tools.
Each agent has a default rung, so a careful agent can start in Ask while a trusted one starts in Auto. A workspace can also set a ceiling that no chat in it can go above, so a shared workspace stays within the bounds you want. You can still turn any single chat down to a more cautious rung whenever you like.
When the agent reaches something that needs your say-so, it pauses and shows a confirmation right where the action would happen, with Allow once and Deny. Approve it and the agent continues. Deny it and the agent moves on without doing it.
When several actions are waiting, a compact banner appears above the message box. From there you can review them from the top, allow a whole group at once, or deny them all.
In Ask mode the agent checks in often, which is a good way to watch a new agent work before you trust it with Auto.
Press Stop to end a run immediately. Anything the agent already did stays done, and anything waiting for your approval is cancelled. You are always one click from stopping, at any autonomy rung.
When an agent hands part of a task to a worker, one dial governs the whole tree. If a worker reaches something that needs approval, that confirmation surfaces in your chat, so you never lose sight of what is happening a level down. See Subagent conversations.
Related articles
Agent permissions and tools
Control what your agents can reach and which tools they can use.
Agents
Create AI personas with custom instructions, their own memory, and their own default model.
Chat settings and toggles
Customize each conversation from the composer: tool toggles, autonomy, compaction, and a spend limit.
Subagent conversations
Let agents delegate parts of a complex request to worker conversations.
Was this helpful?