Imagining Teaching with AI Agents . . .

Teaching with AI is only one step toward educational change, what's next?

More than two years ago I started teaching with AI in my classes. At first I taught against AI, then I taught with AI, and now I am moving into unknown territory: agents. I played with Manus and n8n and some other agents, but I really never got excited about them. They seemed more trouble than they were worth. It seemed they were no more than an AI taskbot overseeing some other AI bots, and that they weren’t truly collaborating. Now, I’m looking at Perplexity’s Comet browser and their AI agent and I’m starting to get ideas for what the future of education might hold.

I have written several times about the dangers of AI agents and how they fundamentally challenge our systems, especially online education. I know there is no way that we can effectively stop them--maybe slow them a little, but definitely not stop them. I am already seeing calls to block and ban agents--just like I saw (and still see) calls to block and ban AI--but the truth is they are the future of work and, therefore, the future of education.

So, yes! This is my next challenge: teaching with AI agents. I want to explore this idea, and as I started thinking about it, I got more and more excited. But let me back up a bit. What is an agent and how is it different than Generative AI or a bot?

Generative AI = Prompt Driven

To use AI, you need a prompt. The prompt might be something simple like “Give me some ideas for what I should make for dinner tonight.” That is known as a “zero-shot prompt,” because you don’t give any examples, you just come out of the blue and ask the Generative AI for something--information, a picture, a video, etc. If you want to get fancy, you could upload a picture of what is in your refrigerator or give a list of the ingredients you have, then you could ask a more contextualized question. That is a few-shot or multi-shot prompt. Those type of prompts contextualize the information so that the generative AI can more easily come to the answer that you want.

If you want to get really fancy, you can start using “Chain of Thought” prompting, where you break down a task into smaller steps and ask the Generative AI to do something more complex. I use this type of prompting in my Socratic tutoring prompts. Some alternatives to this include asking the Generative AI to generate some knowledge about a topic, then use that knowledge to complete a task, or compare information and present it in a table, or check its answers to make sure there is consensus for its answers to a problem. These are all different types of “Chain of Thought” prompting.

As prompts get longer and more involved, a lot of prompting is hidden inside “bots” or “GPTs” where complex prompts can be stored and called upon easily by users. This is a sort of mini-program that runs very complex and sometimes multi-step prompts to run AI systems.

Agents = Goal Driven

An agent is related to simple AI systems in the use of large language models, but it is different in the way that it works in the world. AI agents are autonomous software systems that perceive their environment, reason, make decisions, and take actions to achieve goals. Where generative AI is prompt driven, Agents are goal driven. You tell them what you want, and they figure out the rest.

This is how agents work: let’s say you built a bunch of AI bots with complex prompts. Then, you want them to work together in a system. You would make a “controller bot” that orchestrates and plans, then a collection of “tool bots” like web search, code execution, memory, etc. Those bots would work together as one autonomous unit-- receiving input, analyzing it, perceiving the context, deciding the best course of action, and executing that action through a response or a change to their digital or physical environment. Then, the system would analyze the outcome of the action and learn from the experience, integrating that knowledge into the next loop of learning, and even collaborating with other agents to complete complex tasks in the best and most efficient way.

Agents are a lot easier to use, and a lot harder to control. Because they are autonomous, they can make errors, then compound those errors quickly, so it is important to keep a close eye on what is happening with an agent that you let loose in the wild. That’s the scary downside.

The wonderful upside is that they can also work really well, learn from their errors, and compound their knowledge quickly to be even more helpful. That’s the part I am excited about. I am thinking about how they might completely change everything we know about education, about teaching, about how we run our schools and institutions. Here’s some ways that I am thinking that we can start to use agents in teaching—now, and maybe a few years in the future. Strap in. This might be a wild ride!

Agents as TAs

Right now, I can’t do this because I don’t want to violate the privacy of my students or the integrity of the Learning Management System (LLM) at my college, so I can’t use agents this way, yet. However, I am envisioning a moment in time where I can employ an agent as my teaching assistant—helping my students understand how to effectively use the LMS, checking over my class to make sure I opened the right module at the right time, explaining the syllabus, the grading programs, and nudging students who aren’t coming to class or finishing their work. I imagine giving an agent the following goal:

Every week check to see which students haven’t been coming to class, haven’t signed into the LMS, and haven’t completed their work and ask them if they need assistance. Send me a report on Monday morning at 8 a.m.

Or maybe this:

Make me a list of assignments I need to grade. Coordinate it with my calendar, and make me a grading schedule that takes into account my need to eat, sleep, and get up once in a while. Pay attention to which type of assignments take longer to grade so you can adjust the schedule as the semester progresses. When I open my LMS, greet me and open the next lesson to grade. Encourage me to finish when I start to lag so I can get them done in the allotted time.

Wow. I could also have it check over my class to make sure I am scaffolding my lessons correctly, and ask me if I want to change things . . . what about warning me if I am asking my students to do too much, or too little? It could keep track of how and when my students turn in work to help me gage the workload and due dates. It could show me which assignments were more effective, and which I need to work on—and even survey my students periodically to keep me in the know.

Oh my! An agent could also keep track of the learning outcomes for my department, division, college, and university and tell me how each was or was not fulfilling those learning goals. It could tutor me in better assignment delivery systems, better learning designs, better quizzes and discussions as I created my class. It could automatically adjust a new class I was adding, change the dates, adjust the scoring, and take into account weird calendar situations—like my Jewish holidays.

Sigh. I wish this was possible. The agents are capable right now, but they are also a leaking sieve of information grabbing, privacy evading, money-sucking turncoats who have no allegiance to me, my college, or the laws of the United States Department of Education. I can’t trust them. Yet. But there will be a time, people. A golden time . . .

Agents as Educational Doulas

(OOH! I think I just coined a term!!)

Imagine a time when we don’t have any instructional design, no textbooks, no learning management systems. Imagine a student sits down at a computer, or simply sits on a bench in a park wearing some personal AI device, and says to an AI agent, “I want to learn to write.” Then the AI agent looks at the output of that person’s writing so far, or gives them a diagnostic test, and then pulls up information from an OER database that is pertinent to their understanding at this point. The student learns a lesson, then the Agent adjusts the next lesson to challenge them. Meanwhile, the agent contacts a college, enrolls the student in the correct level of classes for the student’s skill, and pings me. The student joins my class—either in person or online—and I teach them at their level.

Meanwhile, the agent keeps an eye on this student and the other students in the class, encouraging them to attend, tutoring them, explaining things they don’t understand, and reporting to me so I know how best to help that student achieve mastery of the information. Did they do their reading? Did they finish the draft? What areas are they most struggling to understand? The goal won’t be grades, it will be knowledge and skill. Once they achieve mastery, they move on to another class, another instructor, another experience. Like a video game, they repeat things until they understand them, level up when they know what they need to know, and consult with me along the way because humans react to and learn from humans better than any machine. I know how to challenge them the best, give them empathy when they need it, and when to flash the “mom look.”

Classes don’t have a “start” and an “end” date. They don’t go for a particular number of days. It’s all about mastery, learning, succeeding in a goal. It is available when it is needed and for as long as it is needed, adjusted to the developmental and educational level of the student automatically, with a personal educational doula. Classes are still taught because the social aspect of school is still important to human relationships and human experience, but the students may change weekly, picking up what they need and moving on.

This is my vision. Who knows how much will come true, or how long it might take. Meanwhile, on the way, I will work to integrate agents as much and as often as I can—as soon as they are more secure, more private, and approved by my university. Until then, I will sit and dream of all the things I will do . . .