For the last few months I've been going deep on agentic AI. Building them, breaking them, figuring out what actually works.
In that time, I've noticed the appetite for agents has jumped - among clients, peers, people I talk to online. Everyone wants them. And over the last ten days, I responded to that in the most direct way I could think of: I set up ten of them. AI assistants for professionals. Claude-managed agents scoped to specific business tasks. And a few personal ones for myself - Nora, who helps me track calories, workouts, water intake, and mood. BayMax, who handles my scheduling, to-do lists, and helps me stay on top of project management for NeueStudio clients.
The ability to deploy agents effectively is still being figured out. It's uncharted enough that nobody quite agrees on anything yet. There's a lot of room for learning and debate here.
I wanted to document what I'm learning as I go. And the first thing I keep coming back to is scope.
When you first set up an agent, there's a strong pull to give it access to everything. Every tool. Every piece of data. Every API key you can think of. In theory, you're creating this all-knowing, read-and-write-to-everything kind of assistant. It sounds powerful.
In practice, more access doesn't mean more useful.
What I've found - and this lines up with what Dan Shipper over at Every has written about too - is that agents work much better as specialists than generalists. When you give an agent a hundred tools and a hundred thousand words of context, hallucinations go up, not down. Its ability to prioritise and make sense of information falls apart without a specific job to focus on. And if you're running on a token-by-token basis, the cost of hauling that much context into every prompt gets expensive fast.
With Nora, I took a different approach.
Nora has access to exactly two tools: Telegram, so she can talk to me, and a Google Sheet. That's it. One job: keep track of my nutrition and fitness, and keep me accountable to the goals I've set. She doesn't need my calendar. She doesn't need my email contacts. She doesn't know anything about my work.
And honestly - an AI agent plus a Google Sheet can do a surprisingly large amount with just those two things.
Nora is, so far, the most effective agent I've built. Albeit her job is simple (which helps). But because of that specific, narrow context, she just works. There's no confusion about what she's supposed to do, she doesn't hallucinate and she runs efficiently.
This principle isn't new. Plato wrote in The Republic that “the beginning is the most important part of the work.” Seneca put it even more plainly: “If one does not know to which port one is sailing, no wind is favourable.” People have been saying versions of this for a few thousand years. It's just that now the thing you're trying to steer is an AI agent, and the stakes of not knowing your destination are measured in hallucinations and API costs.
Thus the question - one that I hate facing, honestly - is: what do I actually want? That might be the most important question you can ask when building an agent. Not “what can it do?” but “what do I actually want it to do?”
P.S. This week I helped a client set up and learn the ropes of Claude Code. If you'd like to book in a working session - get in touch.
P.P.S. I came across this episode with Dan Shipper and Andrew Wilkinson talking about personal AI use - featuring an AI that helps you choose your outfit in the morning, reminds you of your kids' school events, manages your relationships. Worth a listen.