Compliance-Ready
Agents ask
humans decide.
Keep sensitive decisions with your team, with approval requests and a record of what was reviewed.
Create pull request
Branch: feature/add-logging. Target: main. 3 commits, 127 lines.
Review before action
Destructive operations require human approval. Durable, one-shot, audited.
One-shot capability tokens
Approval is not a suggestion. Server generates a signed token valid for exactly one operation. Token is scoped to org, user, run, and the exact operation. No replay, no guessing.
Durable audit trail
Every approval/denial is logged as a durable event. Reload the session weeks later. You'll see who approved what and when. Essential for compliance.
Gated refusals name their remedies
If an approval is denied, the agent sees the exact reason: 'GitHub write permissions not granted' or 'Automation approval required'. Agent can self-correct, not hallucinate.
SEE IT WORK
One agent. Real tasks.
UseAgentExample workflowOur widget service returns 500 on DELETE /widgets/:id when the widget has 3 or more linked components. Fix it.
Clone and search
Cloned repo, found DELETE handler in handlers/widgets.ts. Searched error logs. Every 500 has 'permission check failed' after component count check.
Reproduce locally
Set up test DB with 3 linked components, sent DELETE request, confirmed 500. Traced the stack: permission middleware runs after cascade check, causing the 500.
Fix the order
Moved permission check before cascade. Added test case: DELETE with 3+ components now returns 204. Old test suite still green.
Open PR
Created PR #487: 'Fix widget delete with linked components'. Added test case to prevent regression.
Questions, answered.
What operations require approval?
Anything destructive or privileged: GitHub writes (create PR, merge, comment), automation delete, subscription changes. Knowledge write is auto-gated (human review in the learning lane). Reads are never gated.
How long does an approval last?
One operation, immediately. After you approve, the token expires. If the agent needs to do the same operation again (next turn), it asks again. This keeps each decision human.
Can I pre-approve certain operations?
Not yet. Each approval is human-reviewed in the moment. Future versions may allow org-level policies (e.g., 'auto-approve PRs under 50 lines'), but shipped: every gate is human.
Start with one task.
See what you can hand off.
Explore a completed run, or join the early-access list to hear about getting your team started.