Everyone sells you the install. Nobody tells you what happens in month three. This is the other half, written from 1,388 logged production failures with the fix attached to each one.
There are a hundred guides on getting Claude Code, OpenClaw, Codex or Cursor running. They are mostly good. They mostly work. You followed one, the thing started, an agent answered you, and it felt like the future arrived.
Month three is when your agent stops remembering what it decided in month two. It is when a scheduled job that ran fine for eleven weeks quietly stops and nothing tells you. It is when a routine update reverts a fix you made in April and the tool keeps running, so nothing looks broken and the capability just disappears. It is when you open a bill and find that ninety percent of it went to a frontier model summarising text a free local model could have handled.
None of that is in the install guide, because none of it happens during the install.
Before you buy anything, run the check. One bash file, twenty-one checks, offline and read-only. It tells you which of the failures in this manual are already sitting on your machine.
$ curl -fsSL https://raw.githubusercontent.com/belkenbot/agentcheck/main/agentcheck -o agentcheck $ chmod +x agentcheck && ./agentcheck agentcheck v1.0 - what is going to break your agent setup read-only, offline, nothing leaves this machine [ CRITICAL ] Blanket wildcard in your permission allowlist why You turned off the one thing standing between a plausible-looking plan and rm -rf on the wrong path. [ WARNING ] You have hand-patched an installed package why The next update overwrites your patch and takes the behaviour with it. Silently. The tool still runs. [ all clear ] nothing to fix here
No network calls. It never writes, moves or deletes anything. When it finds an exposed credential it tells you the file and the permissions, never the value. Read it before you run it. It is one commented file.
Each part opens with what actually broke, explains why the obvious fix does not hold, then gives the version that survives contact with a real fleet. Every part ends in a checklist you can run down in ten minutes.
Why your agent forgets, what a durable memory layer actually looks like, and the silent truncation that stops the bottom of your instruction file from ever reaching the model.
Where the bill actually goes. Routing the cheap work off the frontier model without losing anything that matters, and how to tell which is which.
What your agent can do that you have not thought about. Permission design, unattended execution, and where scheduled jobs are allowed to land.
Credentials in config you forgot was world-readable, .env files in repos that do not ignore them, and why any text passing through a shell is an injection surface.
What an update or a reboot takes out. Hand-patched vendor code, version dead zones, backups of the config you spent months tuning.
Would you know if it stopped? Doorbells, error paths that are not /dev/null, and the difference between a log and a signal.
Every check and every chapter traces to something that actually happened on a four-machine fleet running real production work since March.
A scheduled job configured against a live session key injected a foreign model into that conversation, forced a compaction and wiped working memory. The symptom showed up days later as "the agent forgot everything" and took weeks to trace.
A hand-patched dependency was reverted by a routine update. The tool still ran, so nothing looked broken, and the capability it provided quietly disappeared.
One hundred and eighteen connection-refused lines in a log were written off as cosmetic noise. They were the actual blocker. Nothing shipped for weeks because the real error was hiding in noise nobody had cleaned up.
A credential was hardcoded into a scheduled job's prompt. After a rotation the job read the current value and put it in an alert message.
You already have agents running against real work. Claude Code, OpenClaw, Codex, Cursor, or some combination of them. It is past the demo stage and something you care about depends on it staying up.
If you are still picking a tool, this will not help you yet. There is no prompt library in here and nothing about agents replacing your team.