I’m constantly bemused by the human capacity to throw around words confidently when we don’t fully understand what they mean. LinkedIn particularly invites this. Half the posts read like the 2026 version of the Shakespeare Insult Generator. Call it the “multi-agentic LinkedIn post generator.”
Let’s pause for a moment and consider the analogy in that joke before moving on, because it’s actually helpful for thinking about AI in general and agentic AI in particular. The Shakespeare Insult Generator has a stance and skills. It constructs insults the way Shakespeare does. It knows Shakespearian and Elizabethan insult words. If we were creating that software today, we could write it in one line: “Act like a Shakespeare insult generator.” If you’ve learned any prompt engineering skills at all, you’ve probably learned “Act like a…” or “You are a…” prompt structures. The rest is scaffolding. “Write vivid one-sentence insults, such as…”.
The term “multi-agentic AI” is being thrown around as if it’s magic. It isn’t. If you understand how to create a Shakespeare insult generator prompt, you probably know enough to grasp the basics of the concepts. I’ll walk through those concepts in this post, but first, here’s the TL;DR on why you should care. One one hand, multi-agentic AI fits the mental model of academic work because it’s human-centric. “Act like a Learning Designer who specializes in backward design and writing learning objectives.” On the other hand, multi-agentic AI can have explosive variable cost, for reasons I’ll explain. As I discussed in my last post, the need to control variable cost is a big problem for widespread adoption of AI in educational institutions any time real soon. (Adoption by individuals with consumer subscriptions is another story.) Multi-agentic AI increases that problem multiplicatively. It has its place in today’s EdTech, but it’s not the Next Big Thing that all the hype-y LinkedIn posts would have you believe. It’s probably still a few Next Big Things away for education.
Understanding agents
You’re almost certainly using AI agents now. If Zoom automatically summarizes your calls, Gmail summarizes your email threads, or Claude generates a daily summary of tasks and appointments, that’s agentic AI. A trigger other than a prompt initiated the AI’s action.
I’m going to take a very short detour into language here, because I think it helps make confusing AI talk more legible. The word of the day is “polysemy:” The ability of a word to have more than one meaning. Sometimes those meanings don’t overlap. Nobody will confuse the bank that is the side of a river with the bank that is the place you put your money. Sometimes the overlap is more interesting, “The good book that got destroyed in the rain” contains two meanings of “book” at the same time. There’s the collection of words that could be printed or displayed anywhere while being the same “book”, and there’s the physical object. So “the good book that got destroyed in the rain” is really one physical instance of a long-form writing piece. AI language is rife with interesting polysemy that can be both helpful and confusing. We can say that an AI is trained, and that it performs inference, and it is prompted. We probably wouldn’t say, “Based on on its training, your question prompted the AI to infer the answer.” Whenever you see a polysemous AI word, play with it. You’ll learn where it carries weight and where it obscures differences.
“Agent” is one of those words. You can have a travel agent, a real estate agent, or a talent agent. Those agents may or may not be AI agents. They act on your behalf. Do they have agency? Wrong question. The right question is “Where do we want them to have agency?” (Notice whether the meaning of “agency” changes for you from the first question to the second.) At Learning Impact, I heard a great line—which iDesign’s Whitney Kilgore introduced to the conference—repeated multiple times: “We don’t want humans in the loop. We want humans in the lead.” AI should reduce non-cognitive work and cognitive load but not actual human thinking. Ethan Mollick has used a term coming out of Wharton that I like: “cognitive surrender.” In education, if the AI helps you improve your thinking, that’s good. If it helps you reduce the kind of cognitive exercise that education is about, that’s bad. That’s as true for the educators as role models as it is for the students.
It’s a very different frame from the way the AI industry is promoting agents in software development and some other “coworking” tasks. “Build me a time management app that runs on an iPhone. It should meet needs X, Y, and Z, and have features A, B, and C. Once you’ve finished, make it twice as awesome. Ima go watch TV; text me when it’s done so I can try it out.”
“Human in the lead” is a claim about desired agency, which is a also claim about the degree and kind of automation that the education sector wants right now. As we’ll see, that line has implications for both plausible benefits and cost risks. What do you want to automate in course design? In teaching? In credit assignment? Career pathing? Of course, automation is not a neutral word either. It has industrialization connation. “Personalization”, in the wider world outside of EdTech has arguably degenerated into double-speak for “ineffective automation plus corporate surveillance.” That’s a deliberately harsh characterization, but it’s not untrue. AIs—and by extension, agents—can do something different from the previous generation of “personalization”. But I’m not sure if that buys us much in terms of “humans in the lead vs. in the loop”. Labels alone do not solve our problem. We have to decide what we trust and want AI to automate in education. Those are two separate questions. “Coworking” is weird as a verb in this context, even though the polysemy would technically allow it. “I coworked with Claude.”
Yeah, that sounds weird.
Anyway, “human in the lead” gives us a framework for thinking about AI roles, at least in work that involved direct engagement with the student, curriculum planning, and anything that carries weight in the core educational processes. Putting this together, an agent is an AI that has a role and does work automatically without having to be prompted every time. For example, every time you drop a learning design document into a folder, the agent might (for example) generate candidate assessment ideas.
Let’s talk about cost, since understanding that now will help in the next section. On pay-per-usage AI contracts (as opposed to the flat-fee app subscriptions that most consumers use), the companies charge for two costs: input and output. How much data are you feeding the model, and how much is the model spitting back out? But those two costs bury two other costs: thinking (or “reasoning”, in AI-polysemy) and knowledge work (or “tool calling”). If you translate this into human terms, you won’t go too far wrong. First, you give the person the context they need to do the job. Then they think about the work and do it. Finally, they write up or explain the work to somebody who needs to do something else with it. Each of those steps has a cost. With AI, that cost is in money and electricity.
The promise of multi-agentic AI
Once you have the idea of a persona with a set of skills and rules that automatically follows through on a sequence of actions, you can develop intuitions around multi-agentic AI. Imagine a team of learning designers, each of which has particular skills and is assigned to perform a particular task on the way to designing a course. Imagine also that they have workflow and communications tools so that one designer can know when there’s work waiting from another designer.
That’s it. That’s multi-agentic AI in a nutshell. One agent picks up a task, gathers the context, thinks about the task, does the work, and passes the result on to the next agent, which then performs its piece of work.
There might be still other agents in the mix that, for example, check quality to make sure the work meets quality standards and can push that work back to another agent if necessary. This sort of “checker agent” pattern is most important for elaborate workflows where quality really matters. The good news is that, in a “humans in the lead” scenario, perfect quality may not be necessary or even desirable. The goal is to maximize the amount of time the human in the lead is spent exercising expert judgment, as opposed to grinding through the drafting work. Cognitive support, not cognitive surrender. While better quality from the AI can help sometimes, a human-in-the-lead goal is much more tolerant of some kinds of limitations. A system that enforces particular notions of polish and completeness in automated AI output can nudge humans from leading the design work to providing a stamp of approval to an entirely automated process. This isn’t new to EdTech; auto-graded homework systems and plagiarism detectors are two examples where there is genuine, nuanced, and one-size-doesn’t-fit-all debate about how much automation and seemingly “finished” work for the instructor is good.
…and then there’s the cost
Now for the bad news: Every agent in the chain is eating and spewing tokens everywhere. Each one gathers context, reasons, calls tools, and passes context onto the next agent.
Boom. Cost variability explosion.
The variability problem can be amplified in a nasty way because of a simple but fundamental problem: Context is lossy. Each agent is telling the next what it thinks the other one needs to know. Think about how this works with humans. Sometimes you don’t share a detail because you don’t know that it matters for the next person’s task. Sometimes you misunderstand a detail that’s shared with you. We have a term for this: It’s called the “game of telephone.” Once the context chain is corrupted, it’s hard to audit and figure out where it went wrong. When you add that opaque error rate into the speed at which AI produces outputs, you end up with a system that could easily and unpredictably cost a lot of money to make and keep reliable.
Cost can be managed in some areas, particularly if the multi-agentic system is relatively simple and well-designed. For example, the aforementioned iDesign sells a multi-agentic learning design tool called Align. For those not familiar with the company, they provide fee-for-service Online Program Experience (OPX) offering as a cost-effective alternative to Online Program Management (OPM). They provide learning design, practicum management, and other services. OPMs, from which OPXs evolved as a product category, are companies that provide up-front cash in return for revenue sharing over a period of years. OPXs like iDesign are similar in that they are treated and priced as strategic multi-year investments in building the institutional capacity to serve students and generate sustaining revenues. Institutions invest in and build online programs for the long haul. OPXs are different in that they don’t front the money and don’t take a revenue share. You get what you pay for, and you pay for what you get. In that kind of a cost structure, charging for a multi-agentic system as part of a larger strategic service contract works. Institutions can budget for that. That’s a very different scenario from “Every teacher gets an AI they can ask to do anything they want with all the course materials they want for a fixed price.” iDesign’s product begins to give us some clues about where early AI wins will be in education. They have to support the cost structure of the product.
It’s not going to be easy to avoid multi-agentic systems and the costs they bring, even if you want to. It will likely become unavoidable across governance boundaries, whether that’s between the registrar and the educator (or in software terms, the SIS and LMS), or between institutions. An agent representing one stakeholder group and following one set of policies may need to talk to another agent living in a different world. That cost will sneak in as agents increasingly manage the business logic within applications. It’s a potential hidden cost. I have not yet seen this kind of cross-boundary multi-agentic AI in the real world yet, but I’m confident I will within the next 24 months.
The bottom line: Be cautious about hype-y EdTech claims about multi-agentic AI
“Multi-agentic” isn’t a magic incantation. If somebody makes a claim about their multi-agentic system, ask them what each agent does, how the agents work together, and how costs are managed. You should be able to follow their answers. If you can’t, walk away. And maybe spend a little less of your valuable attention on conference panels and webinars about multi-agentic EdTech for now. I’ll share some thoughts about the realistic paths for AI in EdTech in an upcoming post. But for now, touch some grass, and keep that Shakespeare Insult Generator tab open if you want to respond to multi-agentic posts that are polluting your LinkedIn feed.
Join the Conversation