Agent modes & personas
The agent is a real tool-calling loop that reads, searches, and edits your codebase and runs whitelisted commands. How much it does without asking you first is controlled by its mode.
Agent modes
| Mode | Behavior | Best for |
|---|---|---|
| Interactive | Every file edit and command is shown for approval before it runs. | Unfamiliar code, high-stakes changes, code review-style sessions. |
| Bypass | File edits apply automatically. Whitelisted commands still run without a prompt; anything not whitelisted still asks. | Routine work where you trust the edit but still want a command safety net. |
| Autopilot | The agent runs a full multi-step task end-to-end with minimal interruption. | Well-scoped tasks you're comfortable delegating fully — e.g. "add tests for this module." |
You can switch modes mid-session from the tool window — there's no need to restart a chat to change how much autonomy the agent has.
Named personas
Type @ in the chat box to invoke a specialist persona — each one carries a tuned prompt and a scoped set of tools for its job:
| Persona | Focus |
|---|---|
@debugger | Root-cause analysis, reading stack traces, inspecting live debugger state. |
@test | Writing and running tests, reading structured pass/fail results. |
@security | Security review of the selected code or change. |
@review | Code review pass — style, correctness, edge cases. |
@docs | Writing or updating documentation and comments. |
@refactor | Restructuring code without changing behavior. |
@modernize | Upgrading patterns, APIs, or language features to current idioms. |
@profiler | Investigating performance issues. |
@git | Git operations — history, blame, commit/branch help. |
@plan | Breaking a larger task down into a step-by-step plan before executing. |
Standard unlocks the full built-in set above. Pro adds the ability to define and save your own custom personas with a tailored prompt and tool scope.
Multi-agent delegation
Mid-turn, the primary agent can hand a focused sub-task off to a specialist persona — for example, delegating root-cause analysis to @debugger while it keeps working on the rest of the change — running several delegated calls in parallel when it helps. Standard allows up to 3 delegated calls per turn (2 in parallel); Pro and Teams raise that to 8 (8 in parallel).
Resilient runs
The agent loop runs server-side, so a dropped connection doesn't lose progress — it resumes mid-turn instead of restarting. Tool execution itself (reading files, applying edits, running commands) still happens locally in your IDE, where your code lives.