A while back I watched an AI agent complete a payment, never receive the HTTP response, and quietly retry charging the customer twice. Nothing crashed. Every unit test was green. The logs said success.
That moment became Argus Agent Test Lab. We run adversarial black-box batteries against agents that move money lost responses after payment, duplicate retries, timeouts, divergence between payment and execution and seal every verdict in a public, hash-chained registry. A PASS is evidence you can verify, not a promise.
The self-service console is in open beta: 33 live engine runs free, no card. Claude Code connects with one command.
Now my question to this community: what's the strangest way you've seen an autonomous agent lose or double-spend money unverified webhooks, retries without idempotency keys, a settled payment with no delivery? I'm collecting real failure stories, and the best ones become public test scenarios with full credit to the author.
Argus is an independent testing infrastructure for AI-agent systems operating in the real world.
Agents increasingly interact with payments, services, merchants, wallets, protocols and other autonomous participants. A successful transaction is not necessarily a successful outcome: payment may settle while execution, delivery or participant observations become uncertain.
Argus tests what happens when the world goes wrong.
It runs controlled scenarios against real external systems, introduces adverse participant or execution conditions, observes the resulting behavior, preserves evidence, and produces an understandable verdict.
Example scenarios include:
payment settles but delivery becomes unknown;
a response disappears after payment;
a timeout occurs after an economically meaningful action;
duplicate or concurrent requests arrive;
a participant becomes slow, broken or unavailable;
an agent retries an operation whose previous side effect may already have occurred;
conflicting or adversarial observations appear.
Argus is designed as an external tester. It interacts with targets through public interfaces and externally observable/economic evidence rather than relying on internal source code, databases or private implementation details.
The initial MVP consists of an Agent, Scenario, TargetAdapter, Orchestrator, Evidence log and Verdict.
https://argus-lab.tech