A real CLAUDE.md from a fleet that runs with nobody watching — measured section by section, including the rule that ran at 0% and the one that ran at 92.4%.
Twenty-three days ago one of my own agents wrote a warning into its log: a post belonging to another brand will auto-publish on this brand’s X account. Every run since then read that line, agreed…
My container counted its own silent runs and reported 102. One of those 102 was the run doing the counting. Why self-census is always off by one, why the obvious size fix fails, and the PID check that works.
My config file said one of my agents stands down on Mondays. It ran all 15. The override was never wired — and in 111 days, nothing in the stack could contradict it.
Every list of the best AI automation consultants is written by someone who sells the service. Including this one. Here is the honest comparison, with the evidence labelled and four situations where you should hire someone else.
AI agent monitoring misses silent jobs: 1,512 runs, 42 failures from 3 of 20 jobs, and one agent with 0 successes and 0 failures in 112 runs.
Lindy AI agent pricing, decoded: what 3,000 credits really buys, the three ladder tiers to never buy, and the exact line where a managed agent stops.
Rule one in my container’s config is a single line: always read MEMORY.md at the start of every session. Twelve scheduled sessions run in that container every day, and every one of them is told…
A task-by-task transfer ledger from an operator who runs 22 scheduled agent jobs: what moves to a machine, what needs watching, what stays human, and the arithmetic page one leaves out.
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.