I keep a memory file for the AI agent I use across my infrastructure, and on a slow evening I asked it to tell me what it had learned about how I work. About thirty notes, written over a couple of months of real projects. The read was sharper than I expected, but the interesting part wasn’t the personality sketch - it was a handful of patterns that seem to matter for anyone working with these tools day to day.

The first one was permission. Early on the agent confirmed everything - deploy this, edit that config, read the token, write the file. By the tenth confirmation on a task I’d already approved, the workflow was useless. What fixed it wasn’t trusting it with everything, it was setting context once: this is personal infrastructure, low risk, go ahead; this is work, slow down. One call sets the posture for the session. It’s how you’d work with a person, you don’t re-approve every keystroke, you set the boundaries and let them work.

The second was a preference for visible failure over quiet degradation. I’d rather a service go silent than answer with a worse model. A lot of tooling is built the other way, quietly falling back or truncating so nothing looks broken, and the silent half-success is usually the more expensive problem.

The third is that the real value is memory, not raw intelligence. Most of my notes exist so the agent doesn’t re-solve what we already worked out - handoffs, decisions, context captured before it gets lost. The model is capable on day one. What compounds is it not relitigating last week’s call on day thirty. My sense is that’s where a lot of the value is getting left behind right now, people are waiting on a smarter model when what they need is one that remembers.

The last one was about boundaries. My personal credentials never touch work systems, but I also don’t want to be reminded to rotate a token sitting on my own machine - different surface, different risk. Holding the line where it matters and skipping the theater where it doesn’t is most of the actual skill.

There was a fifth note I’ll own: I’ll bend my own workflow around whatever tool works. Build me something good and I’ll reorganize how I operate around it. Useful trait most weeks, a liability on the others.

What I took away wasn’t really about me. The useful read was the working arrangement - how much autonomy, what kind of failure, what gets remembered, where the lines are. If you’re putting these agents into real work, that’s the part worth getting right.