Approval gates that tell you exactly what will happen.
Gates aren't yes/no prompts. Each gate shows the requested action, the risk level, the evidence supporting the request, and the consequence of approval - so the decision is informed, not reflexive.
What an approval looks like.
Each gate is a self-contained decision with all the evidence in one place - no clicking around to find context.
- Title — the specific action, not a vague "continue?"
- Agent & task — who is asking, and why
- Diff & tests — the evidence behind the ask
- Risk label — low / medium / high, with reasoning
- Consequence — exactly what changes if approved
- Three outcomes — approve, reject, defer
Merge patch into integration worktree
- Agent
- worker.w1 (Studio Engine)
- Task
- add blue test button to sidebar
- Worktree
- feat-blue-btn-w1
- Diff
- +24 −3 across 2 files
- Tests
- 24/24 passed · 11.2s
- Risk
- Low · no deps, no migrations
Studio merges feat-blue-btn-w1 into the integration worktree. No changes are pushed to your branch.
Studio gates anything that changes your system.
Adding, removing, or upgrading packages always pauses for review.
Anything resembling a schema or data migration stops at a gate.
Merging worker patches and applying to your branch each need an explicit approval.
Touching credential files or env vars is gated by default, every time.
Approve, reject, or defer.
"Defer" is its own outcome - it pauses the run without losing context. You can return hours later and the gate is exactly where you left it, with all the same evidence still attached.
- Approve. The agent proceeds with the exact action shown.
- Reject. The agent backs off; the run records the rejection reason.
- Defer. The run pauses cleanly. Resume any time.
Keep your hand on the throttle.
Agent Studio is in private alpha. Drop your email for a build and setup guide.