An agent is mostly the shape of its environment.
After a year of building with language models, the things I tune most are not the prompts or the temperature. They are the tools I let the model see.
Every few years a technology arrives that rearranges our assumptions about what a computer is for. The personal computer was one. The web, another. The smartphone reshaped how we carry the world with us. We are now, I think, living through another such rearrangement — and, as before, the change will be uneven, overhyped for a season, then quietly ubiquitous.
The interesting question isn’t whether this happens. It’s what small practical shifts to make today so the transition delights rather than destabilises us.
The shift from tools to companions
For decades software has been a tool: a hammer we pick up, swing, and put down. The new class of software behaves more like a companion — it has context, it remembers, it anticipates. That change requires a new kind of literacy.
The quiet conversations you have with your tools on a Tuesday afternoon matter more than the product demos you watch on Twitter.
Three habits that have served me well
- Keep a running prompt journal — what worked, what didn’t, why.
- Draft with the model, but think alone. Don’t outsource the hard part.
- Read the boring documentation. The magic is in the margins.
A small example
Here’s the configuration I use for most of my long-running agents. The numbers aren’t sacred — treat them as a starting point:
export const agent = {
model: "claude-haiku-4-5",
temperature: 0.4,
system: readFile("./system.md"),
tools: [search, readFile, writeFile],
stopOnToolError: false,
};The temperature matters less than the system prompt, which matters less than the tools. An agent is mostly the shape of its environment.
What I’m watching next
Local models are getting good enough to matter. The first time I replaced a cloud call with something running on my laptop and noticed no difference in quality, it felt like a small door opening. I think that door opens wider every month for the next year.
Whatever you build with these tools, the most useful thing you can do is keep a careful, honest record of what worked. Write it down somewhere it can be retrieved by your future self at three in the morning, twelve months from now.