It's been a wild start to the year with AI, and it seems the dominant new frontier is agentic AI. AI that isn't just a brain in a jar, but rather has arms and legs to go interact with the world. That idea hooked me, and for the last few months I've gone deep in learning, building and deploying these things into a bunch of different contexts. This has met a surging demand around me for agentic solutions and so in the last 10 days I've built and deployed 10 agents.
Some of these agents are for friends who want to see the hype; some are for directors who want to take organisational load; some are for me and my wife to help us live our lives.
I haven't got my 10,000 hours yet (arguably no one has) but there does seem to be an emergent principle when designing these things that helps them work better, and it's somewhat unintuitive.
I'll start by saying what is intuitive: plugging them into everything. Because why not? More is more? And if it has access to every tool, every email, every word I've written, all my calendar, every document, surely that adds up to a better agent.
Well, at least in my experience, it certainly doesn't.
More access ≠ more useful.
It turns out that the more you give it, and the broader the role you ask it to play, the less effective it becomes. Which, if you think about it, makes sense?
Everything is not parseable and anything is not a job description. What I've seen is that agents set up in this broad way struggle to perform. They don't know what to prioritise out of all the data or what to touch out of the mountain of tools, and so it hallucinates; and it assumes; and it burns tokens trying to figure out a way forward, all while you add even more context through Rage-prompting.
I found this personally with my Openclaw (AI thing) named Baymax. His job was to schedule my day, and read my emails, and write my to-do lists, oh and track my workouts, and can you do my Woolworths shopping, and manage my NeueStudio projects, plus I need you to understand the legislation around AEDs in South Australia + track my sleep, and I'm taking this new B12 vitamin complex can you research it while writing a service agreement for a new client? Baymax tried his best (yes I give my agent pronouns) but he inevitably became confused as I mindlessly shoved in more context and duct-taped on more tools.
So, I started again. This time deciding roles beforehand. I decided I'd have Max, he was the project manager for NeueStudio and would plan my work, and Nora, who would look after my health, tracking my nutrition, workouts, nudging me to drink more water etc. They both have much narrower roles and a specific set of tools to use. Nora, in this case, only has access to Telegram to chat to me and Google Sheets to record. (btw - I am blown away with what you can pull off with an AI agent and Google Sheets). They are both more effective and more efficient while having less information and fewer tools.
Refined inputs = refined outputs
Dan Shipper at Every has stumbled on the same thing. AI agents seem to perform much better as specialists rather than generalists. Just like humans in an organisation, they need a specific role to play.
This principle we are running into is also not a new thing.
“If one does not know to which port one is sailing, no wind is favourable.”
- Seneca, AD 40
And thus like every other design problem, the work must begin with getting clear on what you want to achieve.
- What do I want this agent to specialise in?
- What's the crucial information it needs to know to succeed?
- What are the minimum set of tools it needs to touch to be effective?
- Am I clear on the problem(s) this thing is built to solve?
If I'm a betting man (and I'm not), the organisations and individuals that get real value from agents are the ones who have the discipline to scope upfront, strip context and tools back, and build a specialist.
P.S. This week I helped a client set up and learn the ropes of Claude Code. If you'd like to book a working session, get in touch.