Paper review · November 16, 2026
Primer: an instructor-centered AI framework for building educational simulations
Primer: An AI Framework for Instructor-Centered Educational Simulation Creation
The authors designed a multi-agent pipeline running from instructor interviews to student practice and released an open-source system that lets instructors build adaptive educational simulations suited to their own courses.
Problem
Educational simulations are effective for learning but hard to build. They require scenario design, character development, scaffolding (structures that support learning), technical implementation and repeated revision. Leaving this to a general-purpose LLM often produces the wrong difficulty level or content misaligned with learning goals. The result may also be disconnected from the course context or portray characters in an exaggerated way. Even when a model has pedagogical knowledge, it does not apply it consistently in actual interactions.
Method
The process is split into an authoring stage and a delivery stage, with specialized agents for each step. In the authoring stage, the Intake Agent asks the instructor one question at a time. It collects learning goals, student level, common misconceptions, assessment criteria and course materials, and records a planning journal. Based on this, the Design Agent produces a design document using a Roleplay or Teachback template, which the instructor reviews and approves. The Casting Director then creates character prompts, and the Image Generation Agent creates scene images. The Experience Builder assembles and checks them. In the delivery stage, the Mentor (preparation), Practice Agent (application) and Feedback Agent (reflection) guide the student in turn. Instead of long contexts, agents pass reviewable artifacts such as design documents, transcripts and checklists to one another.
Results
- The authors released an open-source system including the prompts and a GitHub repository. It is not tied to a specific LLM, so local and open-source models can also be used.
- It supports two practice modes. Roleplay targets performance goals that involve dealing with another party, such as persuasion and negotiation; Teachback targets goals where learners check their own understanding by teaching a concept to someone else.
- The authors describe the design document as a boundary object. It is a single document that serves as a lesson plan for the instructor, character information for the Casting Director and a build specification for the Experience Builder.
- The provided text contains no quantitative evaluation of learning outcomes or system quality.
Limitations
The authors state that Primer is an experimental system and should not be used in classrooms without appropriate safeguards and further testing. Currently, variations can be created only within a single interaction; repeated practice across multiple cases is left for future work. Within the scope of the provided text, no user study or evaluation of learning outcomes was found.
Why it matters
It presents a concrete design that combines instructor judgment with the improvisation of generative AI by dividing their roles. For practitioners building educational agents, three points are worth noting: interview-based requirements gathering, a human approval step and artifact-centered handoffs between agents. Because the model can be chosen, it is also easier to meet data governance requirements that differ by institution.
Key terms
- Multi-agent workflow
- A structure in which multiple AI agents with different roles, such as interviewing, design, character creation and running practice, work in stages and pass their outputs to the next stage.
- Educator-in-the-loop
- An approach in which the instructor directly provides information and reviews and approves the design while AI creates content.
- Teachback
- A practice method in which learners explain a concept to a student or peer played by AI. Answering questions and misunderstandings reveals gaps in the learner's own understanding.
- Boundary object
- An artifact that different groups interpret and use for their own purposes while it keeps the same identity. In Primer, the design document plays this role.
Original
Authors: Ethan Mollick, Lilach Mollick, Ben Przystanski, LJ Ciccarelli, Natalie Bach, Laura Zarrow · Posted 2026-11-12 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.