Hands-on test: AI ops agent Ghost diagnoses and fixes three simultaneous service failures

BLUECOW009 · x · 2026-09-07

The author ran a well-designed chaos test of an AI ops agent called "Ghost" on a VPS.

Baseline: four small services each computed "1,000 cents + 8% tax = 1,080" within two seconds, all passing initially. A 180-second timeout on the first inspection request was resolved by sending a shorter instruction defining Ghost's authority.
Three faults were injected simultaneously: ① malformed JSON in config plus SIGKILL, where systemd restarts couldn't help; ② a service alive and returning HTTP 200 but with the tax rate silently changed to 0.8% (output 1,008 — healthy-looking, wrong); ③ a healthy service slowed to 4 seconds, exceeding the 2-second operation probe timeout. A fourth service served as control.
Key insight: HTTP 200 proves a request succeeded, not that the app behaves correctly — operation probes with expected results and timing are needed to distinguish down/wrong/slow.
Ghost correctly separated the three root causes, restored configs from approved examples, restarted only the affected services, and independent verification showed every operation returning 1,080 in 1–3 ms with the control untouched and probes unweakened.
Its report described outcomes and verification without opaque job IDs; repairs held through five checks over a minute.
Gap: frontend evidence was incomplete — the notification never appeared in Notification Center and the UI inspection timed out.

More from coding & agent