Four short notes on what it takes to work with an AI agent over weeks instead of single conversations. Not a framework, not a product — the parts that turned out to matter.
I run a private agent system daily: persistent memory, scheduled jobs, a chat bridge, checks that test the system's own output. Most of what made it usable was not code. It was deciding where truth lives, and building the small rituals that keep it there.
- Files are the memory — a conversation window is a buffer, not a memory. What you want to keep, you write down, in a place that survives the session.
- The close is the write — the important ritual is not the greeting, it's the hand-off. A short, executable next action beats a long summary.
- Verify claims against receipts — an agent saying "done" is a claim. A file, a hash, a test output is a receipt. Build the habit of asking for the second.
- What broke first — the four failures that showed up before any of the interesting ones, and what fixed each.
Anyone who has moved past "prompt it and see" and is now trying to run something that has to be right on Tuesday, with someone else depending on it. It assumes you can read code but does not require any particular tool.
It is deliberately short. Each file is a single sitting, and each one ends with the change to make rather than a theory.
MIT — see LICENSE.