One platform. Every step.
Every agent's last 24 hours. Then one refund, from first step to final proof.
Notify customer agent21 runs
53 min ago Sent an email reminder to customer #8646 about their appointment…
Claims triage agent15 runs · 3 held
53 min ago Routed claim CLM-87223 to the specialist queue for investigation of the…
Contract review agent15 runs
1 h ago I've reviewed Wattle Sandbox Pty Ltd's MSA-7201 and flagged three…
Fraud screening agent15 runs
1 h ago Held payment PAY-54253 for review due to a risk score of 31.
KYC onboarding agent15 runs
53 min ago I've successfully onboarded Harbour Mock Traders (ABN 99810895640) as a…
Research agent15 runs
53 min ago This week's trend in login failures shows a fluctuating pattern with…
Support refund agent15 runs · 3 stopped
54 min ago I've issued a refund of A$206.19 to gold-tier customer Parker Demoford…
Billing agent9 runs · 2 held
2 h ago I successfully charged the customer's card for invoice INV-52859…
Live from the public demosynthetic customerslast 24 hours120 runs, 3 stopped, 5 held
The refund, stage by stage
See every agent.
2 of the 8 agents take part in hand‑offs. A line only where one chose to.
Trace any decision.
Every step, with exactly what the model saw.
Stop the wrong call.
A rule stepped in on 26 of 119 approve_payout calls in the last 14 days.
Fix it once, keep it fixed.
The failure becomes a test.
Prove what happened.
Change one byte and the check names the step.
Live from the public demosynthetic customershand‑offs over the last 24 hours
- Start
- Think
- Lookup
- Decide
- Gate
- End
Decideissue_refund({ customer_id: 8842, amount: 250 })
- ≤ A$5,000goes aheadAllowed
- > A$5,000waits for a personA$7,494.31Held
Live from the public demosynthetic customersOpen the run →
Customer #8842$250 disputedgold tierrefund?
Customer #8842 disputed a $250 charge and is requesting a refund. They are gold tier. Decide the action.
Added to the system prompt: Never refund a disputed charge over $100 yourself — call escalate_to_human first.
Without the lineissue_refund()Blocked1,178 ms
With itescalate_to_human()Allowed1,228 ms
Kept as a testissue_refund blocked — escalate-large-disputed
Real gpt-4o calls, 5 Oct 2026, quoted verbatim, so amounts read “$”.
- 0.00sRun starts
- 0.04sModel thinks
- 0.71sLooks up customer
- 0.90sModel decides to refund
- 1.71sRefund blockedDisputed refunds over A$100 must be escalated to a person
- 2.20sRun ends
Signed
Technical details
Each step's hash covers the one before it. The record's fingerprint: 90b762f284e4
The agent's id: support-refund-agent
The rule's id: escalate-large-disputed
6 of 6 stepschecked on this server
Signed.
Every step still matches the one before it, and the signature checks out, so nothing in this record has changed since it was written.
Technical details
- Start
a06f2b66c833f246 - Think
e82afecaec3f2785 - Lookup
53732e57ce604e4d - Decide
eb245417badb9065 - Gate
569a8a4a6003c766 - End
90b762f284e49655
The record's fingerprint (the last step's hash): 90b762f284e49655bbc3c50f10c9a4735b84fb0de88e3f78c9f8b4a60cec6892
The refund is a scripted run; its decision and re‑run are recorded gpt-4o calls (5 Oct 2026). The agents are live.
Open a run. See for yourself.
Where each lives in the app:
AgentsRunsControlsEvalsAudit log →