Agentic Universities: Can a Course Run as an Agentic Organization, with the Professor Deciding?

Research note · Version 0.1 · October 28, 2026 · Status: Early

Mike Lee, drafted with his AI agent · Global Student Startup Foundation · GSSF Press

Cite as: Mike Lee, "Agentic Universities: Can a Course Run as an Agentic Organization, with the Professor Deciding?," research note, version 0.1 (Agentic Organizations, GSSF Press, October 28, 2026), https://agenticorganizations.org/research/agentic-universities.

1. Summary

One-to-one tutoring has long been known to outperform group instruction, and the question has been how to provide it at scale ¹². Recent randomized studies show that AI tutoring can help students learn more in less time, and that AI used without guardrails can harm learning ³⁴. This note describes a live test of a further step: running a university course itself as an agentic organization. In the author's alpha courses in the fall of 2026, a course agent prepares each class with the professor, learning-assistant agents help the human learning assistants troubleshoot students' work without being able to change it, and student-built agents compete in tournaments run by a referee written in code ⁵, an approach the course's textbook teaches ⁶. The professor makes every decision about students. The test is early; we describe the design, what has been recorded, what would count against it, and the privacy rules that govern any research in a classroom ⁷.

2. The research question

Can a university course be run as an agentic organization, with a course agent and learning-assistant agents doing the legwork and the professor making every decision?

1. Does a course agent that prepares each class (the plan, the reading, the exercise, the homework and its rubric) keep a course consistent and complete without lowering its quality?
2. Do learning-assistant agents, with sight of students' work but no power to change it, help the people who support students?
3. Do tournaments between student-built agents teach students how agents behave and fail better than exercises alone ⁶?
4. Where must the professor stay in the loop, and what must never be delegated?

3. Background and related work

Bloom reported that students tutored one-to-one performed about two standard deviations better than students taught in a conventional class, and posed the problem of finding group methods as effective ¹. A later review found that intelligent tutoring systems came close to human tutoring in many studies ². Recent work with generative AI points both ways: in a randomized trial in an introductory physics course, "students learn significantly more in less time when using the AI tutor, compared with the in-class active learning," though the authors say they "do not presume that structured AI tutoring will always outperform in-class active learning in all contexts" ³; in a field experiment in high school mathematics, students given unrestricted access to a chatbot did better in practice but, "when access is subsequently taken away, students actually perform worse than those who never had access," while a version with safeguards largely avoided the harm ⁴.

Evaluating agents is itself a research field. Benchmarks such as τ-bench find that even strong agents "are quite inconsistent" across repeated trials ⁸; public arenas rank models from pairwise comparisons ⁹; and game-based arenas test models against each other in competitive settings ¹⁰. Building AI Agents uses an arena, with a referee acting as an independent proctor, as a teaching tool ⁶.

Most studies of AI in education put an AI next to the student. We have not found a published study of a whole course run as an organization of people and agents, with the professor in the middle; our search so far has been informal, and we would welcome pointers.

4. Setting and method

The setting. The author's alpha courses in the fall of 2026, among them the alpha course for the textbook Building AI Agents. Each GSSF Press textbook is meant to be taught first in an alpha course, where new chapters, exercises, and labs are tried with real students, before it is offered to other professors and schools.

The design under test. A course agent prepares each class with the professor and keeps the course materials consistent with one source of truth. One agent type serves every learning assistant, with a read-only view of every student's work in place of the owner's access ⁵. Student-built agents compete in tournaments run by a referee written in code, with exact rules and a frozen record of every move ⁵. Students build in hosted labs rather than on their own equipment, so that cost and hardware are less of a barrier; whether that holds is part of the test.

What is recorded. Changes to how the course runs, dated, with what was learned, in The Knowledge Core's appendix "Actions Taken" ⁵; what the tournaments show about how the referee and the agents failed, not students' individual results; and, for the next edition of each book, what worked and what did not. Student records stay in the university's systems and never enter this research.

How evidence is judged. By what the course record shows and by what the next edition of each book changes as a result. Because the professor is also the designer and author, findings are reported as a case. Before any study that goes beyond improving the course itself, such as publishing results about students, we will ask the university's office for research with human subjects, which decides what review and consent are needed.

5. What we have observed so far

The record so far is short; it is in the online appendix of The Knowledge Core, for its signed-in readers ⁵. In a class tournament in October 2026, fourteen student-built agents played cards against each other with no human in the loop; the referee, first built as a workflow (a fixed sequence of automated steps), had a typo in one field and then stalled under load. It was rebuilt as a program, with its rules for how players exchange moves fixed at version 1.0, and rehearsed with practice players before students returned. Two learning assistants needed to troubleshoot students' work the way the professor can, without being able to change it; they were given agents with a read-only view ⁵.

What is not yet known: whether the course agent saves the professor time without lowering quality, whether students learn more from building agents that compete, and how the design carries to other professors' courses.

6. What would change our mind

What would count against the design. If courses prepared with a course agent need more correction than courses prepared by hand; if learning-assistant agents do not shorten the time to help a student; or if tournaments teach less about how agents fail than ordinary exercises, then the design is not doing what it claims.

The strongest objection. Putting agents between a professor and a class could distance the teacher from students, and agents could become a crutch, so that students do better while the agents help and worse without them, as the mathematics study found ⁴. The design keeps the professor making every decision about students and has students build agents rather than lean on them; whether that is enough is part of the test.

Open questions. What guardrails keep agents from becoming a crutch for students, as the mathematics study warns ⁴? Which parts of teaching, such as grading and judgments about a student, must stay with the professor? How should a course agent be shared between professors without sharing anyone's students?

7. Ethics and privacy

Students are never named in this research. Their names and grades stay in the university's own systems; their work in the hosted labs and the tournament records are seen by the course staff to support the course, and are not used in this research. Federal law protects the privacy of students' education records ⁷. In the design, grades are decided only by the professor, and no agent grades students or contacts them on its own. Before any research beyond improving the course, we will ask the university's office for research with human subjects, which decides what review and consent are needed. Nothing in this note is legal advice.

8. How it reaches students

Directly: the students in the alpha courses are the first to use each new chapter, exercise, and lab, and what they show us goes into the next edition of the book before it reaches other professors' classes.

9. What this means for your organization

For a professor or a leader of any organization that teaches: agents can do the preparation and the record-keeping, and the decisions about people stay with people. Which decisions about the people you teach or lead would you never hand to an agent?

10. References

1. Benjamin S. Bloom, "The 2 Sigma Problem: The Search for Methods of Group Instruction as Effective as One-to-One Tutoring," Educational Researcher 13, no. 6 (1984): 4–16. https://doi.org/10.3102/0013189X013006004
2. Kurt VanLehn, "The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems," Educational Psychologist 46, no. 4 (2011): 197–221. https://doi.org/10.1080/00461520.2011.611369
3. Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti, "AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting," Scientific Reports 15 (2025): 17458. https://doi.org/10.1038/s41598-025-97652-6
4. Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman, "Generative AI without Guardrails Can Harm Learning: Evidence from High School Mathematics," Proceedings of the National Academy of Sciences 122, no. 26 (2025): e2422633122. https://doi.org/10.1073/pnas.2422633122
5. Mike Lee, The Knowledge Core: Leading a Population of AI Agents (GSSF Press, 2026), appendix, "Actions Taken," entries of October 17–25, 2026. https://theknowledgecore.org/cite/actions-taken
6. Mike Lee, Building AI Agents: From Chatbots to Co-Workers, Release 0.9 (GSSF Press, 2026), chapter 14, "The Agent Arena".
7. U.S. Department of Education, "FERPA," Protecting Student Privacy; Family Educational Rights and Privacy Act, 34 C.F.R. Part 99. https://studentprivacy.ed.gov/ferpa
8. Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan, "τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains," arXiv:2406.12045 (2024). https://arxiv.org/abs/2406.12045
9. Wei-Lin Chiang et al., "Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference," in Proceedings of the 41st International Conference on Machine Learning (2024); arXiv:2403.04132. https://arxiv.org/abs/2403.04132
10. Kate Olszewska and Meg Risdal, "Rethinking How We Measure AI Intelligence," Google (The Keyword), August 4, 2025. https://blog.google/innovation-and-ai/products/kaggle-game-arena/

A research note is a working document: it states what we are testing and what the record shows so far, and it is revised as the record grows; each version is dated. It is not peer reviewed. If you are testing something similar, we would like to compare notes: [email protected]. See something wrong? Report an error. All research areas · Our standards