Prove your AI agents can still do their job under every rule you gave them. A governed agent fleet can look healthy and do nothing. Every rule is followed. Every change comes back “saved.” Every hourly report is on time. And no work moves.
Rules that cut the path.Over eight days, four well-meant rules in our coordinator’s charter each cut the path from “task finished” to “next task assigned.” Each rule was right on its own.
Compliance with zero output.The day after we fixed those rules, the fleet passed every check and started no new work for 24 hours.
A method, not a rule.The Agent Fleet Test is how we caught both. We run it on our own WWAIOS™ fleet, and we run it on yours.
| Test | What it finds | Where we found it |
|---|---|---|
| 1. Path test | Rules that together leave the agent no legal way to do its core job | Four charter rules blocked task assignment for eight days |
| 2. Receipt separation | Agents logging onto their own instructions until the rules drown | A coordinator page grew to 2.4 million characters, 0.2% of them rules |
| 3. Silence detection | An agent that stops writing and nobody notices | 20 hours of silence, found by a person |
| 4. Prohibition scope | Bans that block internal work along with the risk they target | A “no messages” rule held two internal handoffs for 60 hours |
| 5. Gate evidence | Gates that demand proof from a source that does not exist | A missing webhook stopped every start for four wakes |
| 6. First-receipt proof | Changes reported as done that were only saved, recommended or proposed | Three “done” changes, none working yet |
| 7. Independent review | Agent-written code that ships unchecked | Two high-severity defects caught before merge |
| 8. Claim verification | Status reports that no one checks against evidence | “No bots idle” while two sat quiet for 2 hours 41 minutes |
| 9. Zero-output alarm | A fleet that reports on time every hour and starts no work | 19 clean hourly reports in a row, zero starts, no alarm |
| 10. Authority delivery | Approvals that live in one person’s chat windows and get lost in the relay | Two approvals not found by the agent they were sent to; a placeholder delivered instead of four rules |
| 11. Identity and release | Agent changes that look like a person’s, and merges that reach production, or fail to, without anyone deciding | A merged fix that never deployed; bot work indistinguishable from the owner’s |
The Agent Fleet Test is built into every AI Trust Audit and every Agent Trust Review. It proves your agents can do their job, not only that they follow the rules.
We publish what goes wrong in our own fleet, with the receipts. Agent1 case study · Jimmy Explains
It is built into every AI Trust Audit and every Agent Trust Review. It is the part of the audit that proves your agents can do their job, not only that they follow the rules.
Any fleet that leaves a record of what each agent did. We run it on our own fleet across three AI platforms: two dispatchers and an independent auditor.
Only what we agree in writing. Most of the test works from your agents' rules, configurations and the records they already produce.
WWAIOS™, the WhiteWater AI Operating System. The governance layer that sits around any AI model or agent and checks evidence, authority and receipts before AI-supported work becomes action. Built by Jimmy Standaert and RiGi Group.
