Wharton AI Labs Releases Primer an Agentic System for Classroom Simulations
Primer's Intake Agent interviews an educator one question at a time. It asks about the lesson, the students' level, what they already know, the misconceptions they bring, the teaching approach, and how learning will be assessed. Educators can speak or type their answers and upload materials such as syllabi or rubrics. Wharton Generative AI Labs (GAIL), which publishes the project, says the interview follows backward design, which starts from what students should be able to do and works back to the activities.
The interview produces a design document covering objectives, scenario flow, characters, stakes, and adaptation logic. The educator reviews and approves it, and according to GAIL nothing is built before that approval.
Five builders and three runtime agents
After approval, the work passes through a chain of specialized agents. A Design Agent writes the design document, a Casting Director creates role-specific prompts for the mentor, practice, and feedback characters, an Image Generation Agent produces scene visuals, and an Experience Builder assembles a playable version for one student or a whole class.
Students then work with three agents. The Mentor activates prior knowledge and surfaces misconceptions. The Practice agent plays the student's counterpart in either Roleplay or Teachback mode. The Feedback agent coaches the student from the session transcript and discusses how the lesson transfers beyond the scenario. GAIL describes each agent as both an in-world character and an instructional layer that tracks goals, adjusts difficulty, and gathers evidence. After sessions, educators receive individual and aggregate performance reports.
The three runtime roles resemble the structure of PitchQuest, a venture capital pitching simulator described in a 2024 paper by Ethan Mollick and colleagues at the same lab. That paper built its agents from separate GPT-4 instances, each given goals and system prompts, and it noted the need for rigorous testing.
The reasoning behind the design
GAIL's page says students can often define a concept but struggle to apply it in new settings, and that simulations prompt retrieval, interpretation, action, and revision with feedback tied to what the student did. The lab says it has built simulations for more than a decade and attributes their scarcity to cost and complexity.
The page also states that language models are not pedagogically reliable by default. Left unstructured, it says, they help students too quickly, drift from the learning objectives, and roleplay without educational purpose. Primer's response is to write scaffolding, retrieval, and feedback practices into each agent's instructions and to make the design document and other artifacts inspectable.
What the page does not report
The page gives no empirical results, effect sizes, or sample sizes. It calls Primer an experimental prototype and says it has not been tested as a learning intervention in classrooms or reviewed for FERPA, COPPA, state privacy law, or district policy. Responsibility for that review rests with the institution deploying it.
The code, prompts, templates, and orchestration patterns are released together at github.com/wharton-generative-ai-labs/primer. Primer is model-agnostic, so a school with strict privacy requirements can run it on local or open-source models. GAIL also publishes a prompt that lets an AI agent build its own version of Primer using the full white paper as input, and warns that this can take several hours of agent work and paid model usage. Questions go to GenAILab@wharton.upenn.edu
The project is funded by the Bezos Family Foundation, the Walton Family Foundation, and Charles and Lynn Schusterman Family Philanthropies.
