On November 12, 2026, researchers at the Wharton Generative AI Labs released Primer, an open-source, model-agnostic AI framework that enables educators to author customized, voice-enabled educational simulations directly from lesson plans and rubrics. By pairing an instructor-in-the-loop authoring pipeline with a three-tier runtime system, Primer offers academic institutions a transparent blueprint for deploying adaptive simulations while preserving teacher intent and student data governance.
Wharton Generative AI Labs Releases Primer
On November 12, 2026, an interdisciplinary team from the Wharton Generative AI Labs at the University of Pennsylvania—Ethan Mollick, Lilach Mollick, Ben Przystanski, LJ Ciccarelli, Natalie Bach, and Laura Zarrow—published the preprint Primer: An AI Framework for Instructor-Centered Educational Simulation Creation (arXiv:2611.16986). Alongside the research paper, the team published the complete open-source codebase, prompt templates, and runtime architecture on GitHub (wharton-generative-ai-labs/primer). The initiative was supported by philanthropic funding from the Bezos Family Foundation, the Walton Family Foundation, and Charles and Lynn Schusterman Family Philanthropies.
Educational simulations have long been recognized as high-impact pedagogical tools for professional training, clinical practice, negotiation, and scientific inquiry. By immersing learners in contextual challenges where decisions carry visible consequences, simulations bridge theoretical knowledge and practical execution. However, traditional simulation development has historically presented an insurmountable barrier for most instructors: building realistic scenarios, developing consistent character interactions, scripting scaffolding, and engineering interactive branching logic typically demands specialized software development teams, instructional design consultants, and substantial financial budgets. Primer addresses this bottleneck by using generative AI to lower authoring costs while keeping pedagogical authority firmly in the hands of the instructor.
An Instructor-in-the-Loop Multi-Agent Architecture
Unlike commercial edtech applications that position artificial intelligence as an autonomous tutor dictating curricular content, Primer organizes AI agents around executing the instructor’s specific pedagogical intent. The framework relies on a backward-design paradigm across two distinct phases: authoring and runtime. Rather than asking a single large language model to maintain an entire instructional arc within an unbounded context window, Primer divides responsibilities across specialized agents that exchange structured, inspectable artifacts.
The authoring pipeline begins with an educator-facing Intake Agent. Operating via natural conversational text or real-time voice, the Intake Agent conducts a structured pedagogical interview. It prompts the educator for the target topic, student academic level, prerequisite knowledge, anticipated misconceptions, instructional objectives, grading criteria, and relevant reference materials such as syllabi, lecture notes, or case studies. To avoid cognitive fatigue, the Intake Agent adheres to strict interaction rules: it asks only one targeted question at a time, invites document uploads at any stage, maintains a private planning journal, and conducts targeted research on curriculum standards when appropriate.
Crucially, Primer enforces an approval gate: the authoring pipeline does not construct playable code until the Intake Agent generates a comprehensive design document that the instructor explicitly reviews and approves. Once cleared, backend production agents execute bounded construction tasks: a Design Agent formalizes the scenario structure and adaptation parameters, a Casting Director creates role-specific system prompts and behavioral bounds, and an Experience Builder packages the playable simulation. The resulting output is a portable experience directory containing an experience.json manifest, editable Markdown prompts, scene descriptors, audio assets, and tool definitions.
The Three-Tier Runtime Student Experience
When students enter a Primer simulation, they interact with a three-tier runtime pedagogical architecture designed to balance immersive practice with structured reflection:
- The Mentor Agent (Preparation): Before entering the active simulation, students converse with a mentor agent that establishes the scenario’s context, activates relevant prior knowledge, assesses initial conceptual readiness, and surfaces latent misconceptions.
- The Practice Agent (Application): Students transition into the core experiential exercise, interacting through natural speech or text. Primer supports two core templates: Roleplay, where the practice agent inhabits a counterpart with distinct goals, operational constraints, and conversational stakes; and Teachback, where the agent plays a struggling peer, junior mentee, or advisee requiring guidance, compelling the student to explain and adapt complex concepts.
- The Feedback Agent (Reflection): Immediately following the simulation, a dedicated coaching agent reviews the complete session transcript against the instructor’s rubric, providing constructive, formative feedback tied to the lesson’s learning objectives rather than generic praise.
Throughout runtime, Primer uses low-latency voice pipelines, such as OpenAI’s Realtime API, alongside text alternatives. Student-facing agents maintain a dual role: they inhabit realistic scenario characters while running an underlying instructional monitor that tracks evidence of mastery, calibrates challenge difficulty, and logs observations into a structured session journal for downstream evaluation.
Educator Testing and System Refinements
Between February and June 2026, the Wharton team conducted extensive prototype testing with practicing K–12 educators under an approved institutional review board protocol from the University of Pennsylvania. Participating educators represented both public and private schools across diverse subject areas, including physics, biology, history, technology, and literature. Each educator participated in a two-hour recorded session, first using Primer’s authoring pipeline to generate a customized simulation from their own course materials, and then playing through the resulting simulation from a student’s perspective.
The field trials validated the usability of conversational authoring but also revealed several practical challenges that guided system iterations. Early prototypes suffered from audio turn-taking friction: agents interrupted teachers prematurely during conversational pauses, requiring developers to extend silence detection thresholds before voice models matured. Teachers also experienced navigation uncertainty when confronted with complex multi-step UI checklists; developers replaced these with audible verbal progress markers and clear status indicators showing which backend agent was active. Most importantly, testing revealed that runtime agents frequently provided unearned hints too quickly when students struggled. In response, prompt architects inserted pedagogical guardrails that enforce productive struggle, requiring agents to ask scaffolding questions rather than prematurely surrendering solutions.
Classroom Governance, Privacy, and Local AI Deployment
The Wharton researchers emphasize that Primer is an open research prototype rather than a turn-key commercial product, placing responsibility for compliance and student data privacy on deploying institutions. Using AI in formal educational settings requires careful alignment with the Family Educational Rights and Privacy Act (FERPA), the Children’s Online Privacy Protection Act (COPPA), and state-level student privacy statutes. Instructors are cautioned against uploading student-identifying information, and institutional deployments must establish formal data processing agreements with model providers.
To support varying institutional governance requirements, Primer was engineered to be model-agnostic. While initial demonstrations utilize hosted commercial APIs, the architecture supports deployment with self-hosted open-source models operating on university infrastructure. Running simulations against on-premises inference servers eliminates third-party data transmission risks, ensures full data sovereignty over student transcripts, and provides institutional research offices with predictable operational costs.
Implications for Universities and Research Offices
Primer’s release carries immediate strategic implications for university teaching centers, research computing organizations, and sponsored research administrators:
- Democratizing Simulation-Based Pedagogy: University centers for teaching and learning can empower faculty across humanities, social sciences, and professional schools to build adaptive case studies and clinical interviews without requiring dedicated software engineering grants.
- Architectural Decoupling from Closed Ecosystems: By publishing system prompts, schemas, and orchestration code openly on GitHub, the project provides higher education with an alternative to proprietary vendor lock-in, enabling institutions to audit, adapt, and refine pedagogical logic.
- Rigorous Educational Research Testbeds: For academic researchers investigating human-AI interaction, educational measurement, and multi-agent workflows, Primer establishes an open experimental testbed where prompts, scaffolding levels, and agent configurations can be systematically manipulated under controlled conditions.
- Compliance and Human-Subject Oversight: Research offices evaluating AI-enabled educational interventions can use Primer’s transparent artifact pipeline to document exactly how student data flows between agents, simplifying review protocols for institutional review boards.
Project Availability and Future Research
The complete Primer codebase, prompt library, template definitions, and setup documentation are publicly available on GitHub at github.com/wharton-generative-ai-labs/primer. Additional resources and research updates are hosted on the Wharton Generative AI Labs portal.
The Wharton team noted that while educator testing confirmed the feasibility of instructor-centered authoring, comprehensive empirical studies measuring actual student learning gains against traditional instruction remain the critical next step. Future research will explore multi-student collaborative simulations, automated validation of domain accuracy in generated scenarios, and expanded support for fully local open-weight model architectures under institutional control.
Sources
- arXiv: Primer: An AI Framework for Instructor-Centered Educational Simulation Creation (Published 2026-11-12)
- arXiv HTML: Full Text of arXiv:2611.16986v1 (Accessed 2026-11-15)
- GitHub: Wharton Generative AI Labs Primer Repository (Accessed 2026-11-15)
- Wharton Generative AI Labs Official Portal (Accessed 2026-11-15)
