I grepped my own live stack for hard-coded model IDs. 17 files, 55 occurrences, 6 models — and not one of them is Claude. The brain is the only thing I never pinned.
n8n self-hosted is free the way a puppy is free. The honest cost in dollars and hours after day one, measured against n8n’s own docs and release feed.
A flat 1-credit-per-call API with no way to read your balance for free. Live receipts: three probes, a 970-credit fortnight, and the yield number the meter can’t see.
“AI automation services” means three different purchases: a platform you run, a managed service you rent, or a build you own. Here’s what each actually costs in 2026 — and the one ownership question that sorts every vendor on page one.
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.