I’m running three AI systems every day: Tavi (my OpenClaw instance), Claude Code, and Codex. They can do a lot. And yet the fundamental problem hasn’t changed.
Ludwig von Mises’s framing in Human Action is still the right one: economics is really the study of how humans make choices between options. The binding constraint has always been cognitive load — how much conscious attention you can actually deploy in a day. That constraint doesn’t disappear because you have agents running in the background. If anything, it sharpens.
The classic breakdown: intrinsic load (things you just know how to do — ~10% of capacity), extrinsic load (distractions, obstacles, anything pulling you away from the goal), and germane load (the actual work — focused effort moving you from A to B).
With multiple agents running, the question becomes: where do I put my germane attention? What gets my deliberate focus, and what can I delegate without losing the thread?
This is the thing I keep observing. You can offload execution. You can’t offload taste.
Tavi functions as something like a glorified chief of staff — not because it’s the most capable raw intelligence in the stack, but because it holds the most context. It knows the Second Brain, the projects, the frameworks, the direction. Claude Code and Codex are execution agents. The architecture clarifies when you stop thinking about capability in isolation and start thinking about context.
I use Tiago Forte’s PARA method for the Second Brain: Projects, Areas, Resources, Archive. What I find interesting is that having a shared organizational language with an AI agent creates something like a dialect — shorthand that makes the work faster and more coherent over time.
The dialect point is worth sitting with. Research on linguistics suggests dialects develop fastest in isolated, close-knit groups that communicate frequently under pressure — historically, military units on deployment. Teams develop their own shorthand for the same reason. Something similar is happening with AI agents that carry persistent context.
The larger question I keep circling: what does it mean to build an AI-native organization if you start from the wrong premise? A lot of the current conversation — Tom Blomfield’s Roman Legion framing, the headcount-reduction narrative — treats humans as cost centers. I think that’s a fundamental misread of where value actually comes from.
AI closes the gap from novice to expert. It does not give you taste. Taste is the aggregation of thousands of choices made over time — it’s time-locked, and there’s no shortcut. Companies that are laying people off are, in my reading, declaring that they don’t know how to compound human creativity. It’s the same logic as a dividend: you pay it out when you can’t think of a better use for the capital.
Organizations that are growing headcount alongside AI — like Every — have a different theory. I think they’re right. And the reason maps directly back to Mises: what an organization is really doing is aggregating human choices, and what emerges from that aggregation is what we call judgment, style, taste. You can’t automate your way to that. You can only build the conditions for it.
More on this in upcoming entries.