Essay

Agency, Not Agents

You've already met an AI agent. It drafted a reply to a customer, chased an overdue invoice, or triaged a support ticket before a human ever saw it. Somewhere in your organization, a pilot is probably running one right now. That's agentic AI — and if that's where your understanding of it stops, you're about to file it under the wrong heading: a faster macro, one more automation to fund, someone else's project. The word "agentic" is promising something bigger than that. It's promising agency. And agency is worth five minutes of real thought, because it's the difference between buying a tool and redesigning how your company actually works.

So ask the harder question first: how does something as deep as human agency get built into a machine at all? To answer that, you have to be precise about what agency actually is. It is not intelligence, and it is not speed. It is the freedom to navigate from the discrete to the abstract — to take one specific fact, one specific instruction, one specific moment, and not be boxed in by it. A person with agency can take a rule and recognize when the rule doesn't fit, generalize past the letter of an instruction to its intent, hold many specifics as instances of one deeper pattern. A person without it — however fast, however tireless — is boxed: locked to the specific case in front of them, executing the script correctly and missing that the script no longer applies. That freedom to move off the specific and up into the general is what judgment actually is.

Here's why that question isn't rhetorical anymore. The technology sitting inside your pilots was, quite literally, trained to do exactly that. It's worth noticing what the letters in GPT stand for: Generative Pre-trained Transformer. Pre-trained on an almost unimaginable pile of discrete, specific tokens — this word, then that one, billions of times over. But the training objective doesn't reward memorizing those specific sequences; it punishes a model for staying boxed in them. The only way to get good at predicting the next token across that much variation is to stop pattern-matching the specific and build an internal, abstract representation general enough to generate language it has never seen verbatim — to navigate, not retrieve. That is the same movement, mechanically, as the one that defines human agency: discrete in, abstract structure built, freedom to generate out.

And the pace at which that capability is compounding is worth a beat on its own. GPT-2, in 2019, ran on roughly a billion and a half parameters. GPT-3, a year later, jumped to 175 billion. Frontier systems today train at the trillion-parameter scale, using sparse architectures that activate only a fraction of that capacity per query — more abstraction, at a fraction of the compute cost, built explicitly for the kind of multi-step, tool-using work agentic systems do. Five years in, that curve is still steepening, not flattening.

So don't let "agentic AI" resolve into another item on the automation backlog. What's arriving is a new layer of workforce, sitting between the people who set intent and the software systems that execute rigidly and always have — a layer with a real, if partial, capacity to navigate from the specific to the general the way your best people already do. That's not a task to price out. It's an invitation to think harder than "which process should we automate next," and ask instead what your company could actually become if the system in the middle of it had a measure of agency too. The organizations that sit with that question — rather than skipping straight to the pilot — are the ones who will actually know what they're building.

← All notes