I pointed one client's API token at a different client's queue in production and got a clean 403. The exact schema, the audit that ran before the feature did, and the moment my own tool refused a shortcut I tried to take.
A stale git lock froze the checkout my Cowork tasks run from. Every scheduled agent kept running ten commits behind, reported success, and said nothing. Here is the silent freeze and the sync script that ends it.
By week three the agent forgets every decision you made. It's not the AI — your project has no memory. The files that live in the repo and keep the agent in line, from my real setup.
Adding a new language to 18 articles by hand is a day of copy-translate-paste. I did it in one run: one agent per page, in parallel. The catch is the one file you must never let them touch.
A scheduled agent reads everything I ship and tells me what is worth writing about. I do the writing and the editing. Publishing stays a button only I press.
My notes aren't for rereading. They're the memory my AI works from: a plain-markdown vault Claude Code reads at the start of every session and writes back at the end.
The self-running version of an agent task isn't more code. It's the same task dropped into a scheduled task: a clock fires it, it does the work while you sleep, and a final step delivers the result.
An agent that checks several job boards, drops what doesn't fit, removes duplicates, and leaves you a table of the good ones. You hand it the task once and it does it on its own.
A real website, on a real URL, built in an afternoon with Astro, Claude Code, and Vercel. You don't code it: you describe it and Claude builds it. From zero to online, step by step.
An MCP gives Claude Code new tools from other services (Notion, your database, the web) without you coding the connection. How to connect your first one with one line.
Anthropic's AI agent that lives in your terminal and writes code with you. How to install and start it with one line, plus the two common snags when you begin.