Agent Transaction Sandbox
Describe a policy-gated agent transaction. Watch the sandbox evaluate the policy, build a canonical receipt, and sign it with a separate sandbox Ed25519 key.
Nothing is written to the Agent Commerce ledger, the transparency log, or any other datastore. The sandbox is a teaching tool.
No payment rail is touched. No receipt is stored.
Simulated policy bands
How the sandbox gates transactions
$0 – $500
Allowed autonomously
Under the autonomous-agent limit. Receipt issued immediately; no human-in-the-loop.
$500 – $5,000
Allowed with warnings
Over the autonomous limit. Receipt still issues, but the policy evaluation records a warning that a production deployment would require a spending-plan reference.
> $5,000
Requires human approval
Above the approval threshold. Sandbox still signs a receipt so you can see the shape, but the policy_outcome is `requires_human_approval` and the receipt is not treated as authorization.
These thresholds are deterministic and sandbox-only. Production Agent Commerce policy evaluation is configurable per-agent and lives in a separate code path under apps/api/src/routes/agent-commerce/.
Ready for real agent commerce?
Production Agent Commerce binds receipts to real payment rails, policy registries, and a signed transparency log. This sandbox teaches the mechanics; the live system provides the guarantees.