Previously at the Volcano Base I'd been ruminating about Spring. Since then, I've been finishing off a massive automation project and looking for the next one. It's not exactly tumbleweeds out there but nothing's moving quickly enough to stop me starting to worry a bit. And I finally had to register for VAT, which I've been putting off as long as possible. Bah.

There is a particular kind of digital shame that accumulates quietly in the background of a busy person's life. It looks like a folder. It has subfolders. Some of those subfolders have subfolders. Somewhere in there is an article you clipped in 2019 that you were absolutely going to act on.

You didn't act on it. Nobody does. The article just sits there, a small monument to good intentions, accruing company.

This is what most "second brains" actually are. Not brains. Archives. Glorified filing cabinets with better typography and a suspiciously large font size for the category headings. Read-only memory, in computing terms. You write to it occasionally. You almost never read from it. It exists to make you feel organised rather than to actually help you think.

I know this because I built one. Several, in fact, over the years. They were very tidy. I loved them. But they were completely inert, requiring more maintenance than even I had the patience for.

Something shifted this week. I've been building out an agent team inside the Memex (Volcano Base's operational knowledge base) and the difference between before and after isn't subtle.

The Memex was ROM. Good ROM. Faithful ROM. I'd drop things in: notes, clippings, half-formed ideas. The vault would hold them carefully, like a very attentive butler with a wonky eye.

It's still ROM. That part hasn't changed. But now it's RAM as well.

ROM is storage. It holds things. RAM is working memory: what the processor is actively using right now. The same information, handled in an entirely different way.

Now there's Caldera. An orchestrator that coordinates a small team of specialist agents: researchers, developers, a knowledge architect, a financial director, a marketer, a recruiter. Each one knows their domain. Each one can act on the contents of the vault rather than just observe it.

This week 215 articles were ingested and cross-referenced, a decade of newsletter archives dated and organised, research conducted, topic pages built and updated, connections drawn between things I'd read but never quite joined up. The vault didn't just store the work, it participated in it.

The agents aren't thinking. Let's be precise about that. They're not cogitating, they're not having insights in the small-hours or on-the-toilet sense, they're not lying awake wondering what it all means. They're processing: pattern-matching, retrieving, executing, connecting, at a speed and scale that would be tedious for a human and is apparently fine for software.

But what happens to your thinking when the system around you is doing that work.

I found myself making different decisions this week. Not because I had more information (I've always had too much information) but because the information was connected, surfaced at the right moment, and already partially processed. The cognitive overhead of remembering where things were, and what they related to, simply reduced. That freed something up. Call it attention. Please don't call it bandwidth. Call it the ability to actually think rather than just remember.

The agents don't think for you. They reduce the noise so you can.

Building this required actually deciding what the system was for. Not in a motivational-poster or Simon Sinek sense. In a boring, practical, "what should live where and who's responsible for it" sense.

The agents are only useful because the vault has structure. The structure only works because the workflows are defined. The workflows only exist because someone sat down and thought carefully about the difference between a source document and a knowledge page, between a CRM entry and a project, between capturing something and actually using it.

None of that is glamorous. It's the plumbing. But without it, the agents are just very expensive randomness.

Most organisations building AI systems skip this part. They wire up the tools before they've decided what the tools are supposed to do. They end up with very fast, very confident confusion.

It's not as hard as you might think to create systems like this. Mine is loosely based on Karpathy's LLM Wiki. I added agents wielding skills I'd already created previously. If you have a LLM handy, you can direct it to read Karpathy's notes and have a conversation with you about how it might work for you (or not). Personally, I think this would work better for a very small team or solo practitioner.

Some interesting stuff I discovered on my adventures.

AI's most dangerous predator is a middle manager with a deployment checklist
Ethan Mollick, The Economist

Companies are "de-weirding" AI. Stripping it of its strange edges so it fits neatly into existing processes, and in doing so, are squandering the only thing that makes it worth having. Mollick's three-part model (Leadership, Crowd, Lab) is the most sensible thing written about organisational AI adoption this year, and the observation that employees are quietly hiding their AI use from management is worth sitting with for a while.

Read it (Originally published in The Economist, April 2026)

Japan is automating the jobs no one wanted anyway
Kate Park, TechCrunch

While everyone in the West argues about whether robots are coming for jobs, Japan is deploying physical AI to fill roles that have been sitting empty for years because no human particularly fancied doing them. The framing is refreshingly undramatic: this isn't disruption, it's plumbing.

Read it

The thing you called an "AI agent tool" last year is now just a feature
Andrew Green, n8n blog

The agent landscape shifted while everyone was arguing about prompts. RAG, memory management, web search - these are now table stakes, baked into the major platforms. Green's new evaluation framework pivots from "can it do things?" to "is it enterprise-ready?" Useful if you're choosing tools, or wondering why the one you chose last year feels a bit thin.

Read it