In a hype cycle fascinated by rogue, hacking AI agents, it feels like there’s another kind of agency getting lost in the spectacle, one with specific relevance to higher education. It isn’t absent from discussion: the loss of human agency is a doomer talking point. But human agency often feels disconnected from the agency attributed to machines. The kind of machinic agency a coding agent gains when we let it act autonomously can seem incommensurable or even at odds with the kind of human agency you may or may not have when using one.
Agency, in the sociological sense, is about individual humans having the power and resources to fulfill their potential. Questions about this kind of agency run through poverty traps, credentialism, social capital, face, the economies of time and attention, and so on. It’s far messier than the abstract, philosophical sense of agency: the capacity for an actor to act in an environment. Higher education is rightfully obsessed with (and haunted by) issues of human agency.
Rather than compare and contrast human and machine agency, I want to discuss a secret third thing: transagency, specifically transagentic knowledge acquisition. Don’t bother trying to look it up on Wikipedia. It’s just an idiolectical term, specific to the notes I’ve been keeping on my computer for the last few months. (It has obvious family resemblances to distributed cognition, the extended mind thesis, cyborg theory, and sociomateriality, but I use it to foreground how you, or we, learn in a transagentic context, not just think or exist.) This blog post is a coming-out party for the term because I want to think through the idea more clearly by writing in public.
Imagine you are riding a bike or driving a car, or in any other situation where you’ve wrapped yourself around or through some kind of transportation machine. After a period of clumsy unfamiliarity, when you are first learning to ride or drive, it gets easy to be one with the machine. Regardless of whether it has leather seats, it can fit like a glove. You kind of feel the texture of the road even though the sensation is mediated by the tires, suspension, body, and seat. You feel like your body is taking on a slightly different shape, letting you ignore little bumps in the road but demanding extra vigilance for specific maneuvers like backing up or cornering at speed. You have this new ability to take up space on the road and negotiate position with other riders and drivers in their lanes.
That machine you are driving represents power and resources, and you might use it to fulfill your human potential, or maybe just to arrive at a meeting on time even when you departed late. I’m trying to make it easy for you to agree that machines influence your social agency. The fact that some bikes have electric motors or that some cars can park themselves in a lot without you in the driver’s seat is almost irrelevant. If the bike was actually a horse with feelings and horse-friends, that would still be a distraction. I don’t care much about the machine’s agency, or even the horse’s, for the purposes of this blog post.
I’m interested in this not-entirely-human agency that comes up specifically when a human is driving a rather capable vehicle. Because this is the BayLeaf blog, the vehicle in question is going to be Generative AI stuff, and the drivers are going to be teachers, students, and so on. For me, transagency is this new agency that comes through coupling with the machine. I’m interested in what that means for how transagents learn and gain knowledge through experience in higher education.
Let me offer a story of how I (or more precisely the transagentic-we) learned something recently and how BayLeaf, as a piece of higher education infrastructure, supported it.
I want to make sure people can use BayLeaf services on their personal devices from every common operating system in our community. But I have a laptop running macOS, and I know some folks use Windows. I want to offer a nice gift to my local community that other campuses can copy and remix, but I don’t have a Windows machine handy for testing purposes.
I could open my coding agent (OpenChamber backed by the BayLeaf API), start a conversation rooted in the BayLeaf source code folder, and define a vague task:
Let’s make sure our instructions for running OpenChamber with the BayLeaf API also work for Windows users.
These days, I could probably walk away from my laptop for half an hour and expect to come back to a huge chain of messages ending with something like, “Validated with Windows 11 and OpenChamber 1.24, but it required the user to know the difference between cmd.exe and PowerShell. Do you want me to patch the docs with a clarification, deploy to production, and push the revision back to GitHub?” I could say “yeah,” and it would do it. But I wouldn’t have personally learned from the process, and my coding agent would probably have to stumble through another expensive chain of steps the next time I defined a similar task. Someone, or something, ought to have learned from the experience, not just made the task go away.
In reality, I didn’t do it that way. I sent a message more like this one instead:
I wanna make sure that folks on Windows can follow our instructions for setting up OpenChamber. But there’s a bigger issue: I frequently want to test the Windows-specific instructions I give to my students. I think I saw that Daytona has Windows sandboxes now, so let’s cultivate the skill of cross-platform desktop app usability testing.
Because of agent skills and other context files already present on my laptop, my agent knows I’m a teacher in addition to being the BayLeaf operator. It knows we already use Daytona extensively in BayLeaf operations, so it already has good documentation for how to act as me with my Daytona account. It can find and digest the official Windows Sandboxes documentation on the web. It could have drawn on all of that context to get the task done without me.
What’s important is that I said “let’s cultivate the skill” and named a much broader intent than the initial task. My agent also knows about my idiolectical sense of transagency and what it means to “cultivate” in that context. It knows the difference between an agent skill (the agent working by itself) and a transagentic skill (the agent working with me). So, as it tries to get OpenChamber set up on Windows, it narrates the process in terms of my broader learning goals. It pauses to challenge me with reading-comprehension questions to make sure I’m following along. By the end, we’ve written a lot of text into files outside the BayLeaf repository about how to operate remote Windows machines efficiently. I was highly engaged: thinking, reading, writing. I know how the system works, and the system I have in the end is even smoother than the one implied by the official docs because we’ve built convenience wrappers that shortcut the use cases I care about.
It’s like my car has grown a new wheel or something, but in a useful way, or I’ve grown a new hand. I feel like I’ve got more (social) agency here. I have new power and resources to respond to the human community around me. My agent has perhaps gained some new abilities, but I think it implicitly had those abilities before, just with more stumbling and less efficiency. Importantly, the way I work with my computer has been extended. I shouldn’t use passive voice: I extended it, or maybe transagentic-we extended it. As a result of going on a jaunty drive (it was fun!), the me-and-the-car has extended the abilities of the me-and-the-car.
Now that I have these files on my computer, I could email them to you. You could tell your agent, “Let’s adopt a form of this just for us,” and you-and-the-car could have a time getting it working with your accounts on your operating system. Knowledge here is acquired through experience. It gets crystallized in text, but rather than becoming objective documentation, it is almost subjectified: more diary entry than textbook, suffused with idiolect, made specific to one transagentic coupling. I almost said “individual,” but being able to turn the human/machine analytical boundary on or off is the point.
I can support Windows users without owning a Windows machine, turn a vendor’s generic infrastructure into practices shaped by local needs, and pass those practices to other campuses. Power and resources have shifted, however slightly: I am less constrained by the hardware I own, my community is less constrained by my prior experience, and our future work can begin from knowledge accumulated across me, the agent, the wrappers, and the files. This is how transagentic knowledge acquisition returns to human agency in the sociological sense: cultivating the coupling changes what people can do for and with one another.
Transagency feels highly relevant to and connected with human agency, so folks in higher education ought to know about it. We ought to have accessible, responsibly operated infrastructure to support it. Maybe we ought to teach it in class.