My watchdog reads SKILL_RESULT: success OR EXIT CODE: 0 as a pass. All 42 recorded failures on this container also exit 0 — so it has never caught one. Why a fix that passes its own test isn’t verified.
What is AI automation? A model doing a step in a process that runs without you. Here are the three tiers, and a rule for telling which one your task needs.
One bug, one question, and 16 defensible answers between 12 and 197 — all from the same file on the same morning. The fix costs one log line.
My agent recorded the same four bugs 221 times in 105 days and fixed none of them. The memory worked perfectly. Here is what AI agents actually need in order to learn.
My queue said 107 ideas ready to go. Sorted by date instead of priority, the median was 103 days old and 60 rows had been passed over for three months. The Sunday check that finds a reject pile pretending to be inventory.
A run of mine died after 40 minutes and left a 320-byte log. It had already sent five emails and filed its own success report. Three shortcuts for finding out what a failed AI agent run already did — before you retry it.
I audited 1,399 runs of my own autonomous agent fleet. 49 failures — 48 were one expired token. The real AI agent limitations aren’t about intelligence.
Replicate rents you any AI model by the run, no GPU and no subscription. This morning it charged me double what my own documentation claimed, and the documentation was the thing that was wrong.
A 503 told my agent the vendor was down. It wasn’t. A cached page told me my SEO API was broken for three days. It wasn’t. Your monitor is a system too — and it fails just as silently.
A buyer side guide to hiring an ai automation consultant: what the job actually involves, what it costs, and the five questions that disqualify most of page one.