Logic & Layers
PUBLIC SANDBOX

AI Agent Approval & Audit Starter Kit

Give the model request power.
Keep authority in code.

Try the boundary yourself. The agent can read, policy can deny, a human can approve one exact side effect, budgets can stop work, and retries cannot duplicate execution.

Browser-only sandbox. The shipped starter persists the same workflow through Postgres/Supabase, RLS, approval rows, budget reservations, receipts, and audit events.
01Request

Agent proposes tools.

02Policy

Allow, deny, or gate.

03Human

Approve one exact action.

04Evidence

Receipt + audit explain it.

Run statusready
Pending approvals0
Runs started0
Duplicate effects0

1 · Agent request

Propose a state-changing action

demo.agent

The reference scenario also asks for a safe read and a destructive delete so you can see all three policy outcomes in one run.

2 · Human gate

Approval queue

clear
2
No action is waiting on a human.

Start the governed run above to create an approval.

3 · Evidence

What happened, in order

latest 1
sandbox.readypolicy

Local demo loaded. No shared database or provider APIs are used.

Release evidence

Tested boundary, not a “safe AI” claim.

The locked buyer archive passed 59 unit/component tests, a production build, real Postgres/RLS adversarial assertions, approval/execution concurrency, concurrent budget reservation, SDK integration, browser E2E, and a 390px no-overflow gate.

59 testsreal RLSconcurrencybrowser E2E