Core Concepts
Trust Modes
Three modes. Switched in one click. They control how much autonomy the agent has to act without your approval.
Switching modes#
The active mode is shown in the chat panel. Click it to cycle between Ask, Auto, and Yolo. Each agent profile can have its own default mode configured in Settings. The global mode applies to all agents unless overridden per-profile.
Ask mode#
Every tool call requires your approval before it runs. Nothing happens without you explicitly saying so. The agent proposes each action — read file, run command, write code — and waits. You review what it plans to do and approve or skip.
Good for#
- ◆You want to understand exactly what the agent is doing
- ◆Working in a codebase you have not seen before
- ◆High-stakes operations where mistakes are expensive to reverse
- ◆Learning how the agent reasons about problems
Not ideal for#
- ◆Fast iteration on throwaway code
- ◆Repetitive tasks where every step is obviously safe
Best context: New codebases, unfamiliar territory, learning a project
Auto mode#
Safe, reversible operations run automatically — reading files, running searches, querying the AST index. Destructive or irreversible actions pause and ask: writing files, running shell commands, making git commits. This is the default mode for most developers.
Good for#
- ◆Normal development work where most agent actions are safe
- ◆Production code where you want a check on destructive operations
- ◆Shared repositories where accidental commits would cause problems
Not ideal for#
- ◆When you want to review absolutely everything
- ◆Greenfield throwaway projects where speed matters more than oversight
Best context: Day-to-day work, production code, shared repositories
Yolo mode#
Every tool call is auto-approved. No interruptions, no confirmation prompts. The agent runs at full speed until the task is complete. File writes, shell commands, git operations — all execute without pause.
Good for#
- ◆Greenfield projects where nothing is precious yet
- ◆Throwaway feature branches where reverting is easy
- ◆You have reviewed the task and trust the agent to execute
- ◆Repetitive tasks with a clear known pattern
Not ideal for#
- ◆Production branches
- ◆Any codebase with data you cannot easily restore
- ◆Tasks where the agent's plan is not yet clear to you
Best context: Greenfield projects, throwaway branches, you know what you are doing
Per-agent mode#
Individual agent profiles can have their own mode set in Settings, which overrides the global setting. A useful pattern: Security agent always on Ask (every vulnerability scan step gets reviewed), Researcher on Auto (web searches and reads are low risk), Coder on whichever mode fits your current work.