Intel / article

Primer: Wharton's open multi-agent system for teacher-built simulations

Source: gail.wharton.upenn.edu
Author: Ethan Mollick
Date: Published Nov 12, 2026

Context

Ethan Mollick's Generative AI Lab at Wharton has released Primer, an open-source system in which a teacher describes a lesson in conversation and gets back a playable simulation for students. The lab's starting point is that a capable is not a reliable teacher on its own: left to its defaults, it helps students too quickly, drifts away from the learning objective, and produces convincing roleplay without the structure that makes it educational. Primer's answer is to split the job among several agents, each working from documents a teacher can read, instead of trusting one long chat.

Authoring starts with an interview. An intake asks the teacher, one question at a time, about the lesson, the students' level, prior knowledge and common misconceptions, and how the work should be assessed; teachers can also upload syllabi or rubrics. A design agent turns those answers into a design document covering objectives, scenario, characters, stakes and how the simulation should adapt, and nothing is built until the teacher approves it. Further agents then write the prompts for each character, generate scene images and assemble the playable version. Students meet three agents in sequence: a mentor that activates what they already know, a practice partner that either plays a counterpart with its own goals or asks the student to explain a concept to someone who does not understand it yet, and a feedback coach that discusses specific moments from the transcript. The teacher receives individual and class reports afterward.

The lab is clear that Primer is an experimental prototype. It has not been evaluated as a learning intervention in classrooms or reviewed for student-privacy laws such as FERPA and COPPA, and schools that use it are responsible for that review. Its contribution is the open architecture: the prompts, templates and code are published, it can run on local or open-source models, and the lab also offers a prompt that asks a coding agent to build its own Primer from the accompanying paper, a job it warns could take several hours of agent work and paid model usage. The paper itself could not be retrieved for this explanation, so this account rests on the project page and that build prompt.

Key takeaways · AI-distilled

Wharton's Primer starts with authoring: an Intake Agent interviews the educator one question at a time, a Design Agent turns that into a design document, and nothing is built until the educator approves it.
After approval, a Casting Director writes role prompts, an Image Generation Agent makes scene visuals, and an Experience Builder assembles a playable simulation, each agent working from inspectable artifacts rather than one model's memory.
At runtime, a mentor agent surfaces misconceptions, a practice agent plays a counterpart in Roleplay or Teachback mode, and a feedback agent coaches from the transcript, each with an instructional layer that tracks goals and gathers evidence.
The Wharton team says default models help students too quickly and drift from the objective, so scaffolding, retrieval and feedback rules are written into each agent's instructions rather than left to the model's general knowledge.
Primer is LLM-agnostic and can run on local or open models, but Wharton calls it an experimental prototype not evaluated in classrooms or reviewed for FERPA, COPPA or student-privacy law.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source gail.wharton.upenn.edu