Interactive generator · quest generation
Agentic procedural content generation
Can an AI design a good quest? A role-playing-game quest is a chain of beats — travel, gather, talk, solve a puzzle, fight, a climactic boss, a reward — sometimes with optional side-quests branching off. A good quest isn’t random: it opens gently, builds difficulty to a climax, varies its activities, runs to a satisfying length, and offers a little optional content. That makes “design a quest” an optimization problem with several competing objectives.
Here two agents collaborate. A generator proposes and rewrites candidate quests; a critic scores each one against the design objectives. A genetic loop keeps the best and lets the generator focus its edits on whichever objective is currently weakest — so a random first draft self-assembles into a well-formed quest. Tune the target length, difficulty, and variety, and watch the quality climb.
This interactive is a faithful re-design of the research, simplified for illustrative and teaching purposes. It conveys the idea and behaviour of the model — it is not the paper’s full method, data, or results.
Top: the best quest — numbered beats coloured by type (diamond = boss, square = reward), with optional side quests branching above. Bottom: the generated difficulty arc against the ideal (dashed). Each generation the critic scores the design objectives and the generator rewrites the weakest one, so the quest self-assembles from a random draft.
How the model works
Each candidate quest is a sequence of typed beats, each with a difficulty. The critic agent turns “good design” into five measurable scores: structure (a proper intro, a single boss, a reward at the end), difficulty arc (how closely the beats follow an ideal curve rising to a late climax), variety(few repeated activities, many distinct ones), pacing (matching the target length), and branching(a healthy amount of optional side content). Their weighted sum is the quest’s quality.
The generator agent evolves a population of quests: it clones the best, then hypermutates the clones — but instead of mutating blindly, it reads the critic’s scores and directs its edits at the weakest objective (fixing a broken ending, nudging a beat’s difficulty toward the ideal arc, breaking up a repeated activity). That targeted, feedback-driven rewriting is the “agentic” step, echoing the generate-and-critique loops in the collaborative and generative-AI content-generation work below. Over a few dozen generations the quest converges on a coherent, playable design.
The research behind it
This demo is a teaching stand-in for a body of work on generative and agentic AI for content generation — AI-human collaborative puzzle design, generative models, and LLM-assisted frameworks.
Quests here are illustrative — the demo shows the agentic generate-and-critique loop, not a specific published system. See also the full demo gallery.