Google let an AI agent playtest its own generated classroom simulations before release

Digital learning tools mostly deliver passive content, videos and textbooks, because interactive simulations are expensive to build and rarely match a specific teacher's curriculum or students. Google Research's response, building on its 2024 LearnLM models and a 2025 'Learn Your Way' experiment, was to apply generative UI, letting a model construct an interface dynamically instead of coding it in advance, so teachers can generate custom simulations on demand rather than pick from a fixed library.

The process keeps the teacher in the loop at the start, approving objectives before anything is built, then hands quality control to a pipeline at the end. Each simulation breaks into harder game levels tied to those objectives, with hints, a formula toolbox, and tailored feedback. Before a teacher sees anything, a self-correcting loop checks pedagogy, mechanics, and visual design; the mechanics check is agentic, opening a Chrome instance to interact with the simulation as a user would, including adversarial inputs like an extreme control value, to catch broken levels.

Google released a public library of over 30 vetted learning interactives in English STEM subjects, all teacher-reviewed before publication. In a study with UK STEM teachers, physics and chemistry topics rated best for simulation creation. A separate study with 12 US teachers, each requesting three custom interactives for their own classrooms, produced an average quality rating of 8 out of 10, with teachers noting that dynamic generation solved what static, off-the-shelf simulations couldn't: differentiating instruction to their actual curriculum.

Have an agent try to break the simulation before a teacher sees it

Requiring teacher approval of learning objectives before generation continues keeps a human decision at the one point where a wrong target would make every downstream level wrong too.

Splitting validation into three named criteria, pedagogy, mechanics, and visual design, lets each be checked by whichever method fits it, instead of one generic quality pass.

Using an agentic evaluator that actually drives the simulation in a browser, not just inspecting generated code, catches solvability failures that static analysis would miss.

Testing adversarial inputs, not just the expected path, targets a failure mode a static, hand-built simulation wouldn't have, since it's tested once by its authors and never regenerated.

When you can't afford a human to check every generated artifact, give an agent the same interface a human would use and have it try to break the artifact like a careless or adversarial user would.

Google is piloting the tool with schools through the Google for Education Pilot Program, where newly generated interactives go back to the requesting teacher for review and only enter the public library after that approval, and plans UX research and field studies on learning gains and engagement in real classrooms. The team describes the released library as an early step, with more subjects and grade levels still to come as the pilot expands.