What I learned building a personal agent

I've been building Cael — a personal agent that holds my goals, my reading, my training and my calendar — for a while now. Most of what I expected to matter didn't. Here's what did.

Hosting was the whole problem, until it wasn't

I went through OpenClaw, then Hermes. Every time, I'd get somewhere real and then hit the same wall: infrastructure. Where does it run, what keeps it alive, how do I deploy a change without breaking the thing that texts me in the morning.

Eve removed that question entirely. It's a folder of files that deploys like a Next.js app. Tools go in one directory, schedules in another, channels in another — no registration step, no config file to keep in sync. Drop a file in the right place and it works.

The unlock wasn't a better model. It was that I stopped spending my evenings on plumbing.

Tools compound, they don't add

The first tool is a novelty. The fifth changes what the thing is.

Once the agent could read my calendar and my notes, it started making connections I hadn't asked for — noticing that a commitment I'd written down conflicted with something I'd scheduled, surfacing a thought from three weeks ago because today's task rhymed with it. Every tool you add multiplies against every tool already there.

So the advice is boring and correct: keep adding tools. The value curve is not linear.

Keyword search is a dead end

The first version of memory searched by keyword. It felt fine for about a week, and then it felt useless — because I never remember the words I used, I remember what I meant.

Switching to semantic search fixed it: turn on the vector extension in Postgres, store an embedding for every note as it's saved, and search by meaning. Now I can type "how should I treat people" and get back the right notes even though none of them contain those words.

If you're storing anything a human wrote, do this on day one.

The worklog is the real memory layer

Not the agent's memory — mine, and the build's.

Every session starts by reading a running log of everything built, every decision made, every next step. Without it, each session begins from zero and re-litigates choices that were already settled. With it, work compounds across days the same way tools compound across features.

This is the single highest-leverage file in the project and it's just markdown.

Skills for anything new

Models don't know about libraries released after their cutoff, and they'll confidently guess. Skills — small markdown files that teach the agent a specific tool — fix exactly this. They were the difference between fighting shadcn and Assistant UI and just using them.

Related: before writing any code against an unfamiliar library, have the agent go read the actual repo. The source beats the model's recollection every time.

Get out of the way

The habit that changed my output most: every time I reach for a browser UI to do something, I ask whether the agent could just do it.

Generate an animation. Configure an integration. Design a component. Query the database. The agent reads docs, writes code, runs commands and iterates far faster than I click. My job turned out to be deciding what — not doing it.

The thing that surprised me most is how much it stopped feeling like a chatbot. It texts me in the morning. It remembers what I've been chewing on. It notices patterns I don't. That's a different category of software, and it took a weekend.