Recommended practices
Practical guidance for using AgentDock's real permission controls safely.
AgentDock's safety model is only as good as how deliberately you use it. A few practices that follow directly from the controls covered in this section:
Start conservative on unfamiliar workspaces
When you open a new or unfamiliar project, start the agent in a restricted mode — Claude Code's plan, Codex's read-only, or Antigravity's plan — before granting write access. This lets you see what the agent intends to do before it can touch your files. See Agent permissions for what each mode covers.
Keep destructive-action confirmation on
Settings → Permissions has a confirmDestructiveGitActions toggle that guards file reverts in the Changes drawer. Leaving it enabled costs you one extra click and gives you a chance to catch an accidental revert.
Treat bypass modes deliberately
bypassPermissions (Claude Code), bypass (Codex and Antigravity) skip the interaction-card prompts entirely — the agent proceeds without asking. These modes exist for a reason (fast iteration in a workspace you trust), but they remove your last checkpoint before a change happens. Use them intentionally, on workspaces you trust, not as a default.
AgentDock Pro
AIgency's own run-level bypass permission mode requires a one-time acknowledgment before it can be used, specifically because it's the same kind of trade-off applied to an automated, multi-task run rather than a single conversation. Don't treat that acknowledgment as a formality — see Approvals.
Use the raw terminal drawer when something looks off
If an agent's behavior or a permission prompt doesn't match what you expect, the raw terminal drawer shows you the underlying CLI output directly, which is the most reliable way to see what actually happened.