A training ground to make AI your personal sidekick.
The best ideas don't come from one mind.
They come from many minds, with different perspectives, competing and collaborating.
This pattern shows up everywhere — in AI research, in design practice, and now, in a tool.
Reinforcement learning doesn't need instructions. It needs an environment and a signal.
Richard Sutton & Andrew Barto, Reinforcement Learning: An Introduction, 1998An agent stands in a world with no manual. No instructions. Just a grid of possibilities and the freedom to explore.
Somewhere in the grid is a goal. The agent doesn't know where — but when it stumbles onto it, a signal fires: that was good, do more of that.
The agent wanders, bumps into walls, backtracks. Each failure narrows the search. Each success reinforces a path. Slowly, a policy emerges.
Click on the grid to place rewards and penalties. Watch the agent learn a new policy in real time. You're shaping the environment, not writing instructions.
The deep lesson: you don't teach an RL agent what to do. You design the environment and let the agent discover the answer itself.
A forger and a detective, locked in an arms race that produces masterpieces.
Ian Goodfellow et al., Generative Adversarial Networks, 2014In 2014, Ian Goodfellow proposed a deceptively simple idea: pit two neural networks against each other. A Generator that creates, and a Discriminator that judges. They start knowing nothing.
The Generator produces crude attempts — rough shapes, wrong proportions. Like a first draft with no revision.
The Discriminator spots every flaw. "Too simple. Not realistic. Missing detail." Its criticism is the Generator's learning signal.
Each gets better. The Generator refines. The Discriminator sharpens. Quality spirals upward through competition — not collaboration.
Neither network could improve alone. It's the tension between them that drives quality. Sound familiar? Think: design critique.
OpenAI Gym gave every researcher the same standardized environments to test their ideas. A gym for algorithms.
Greg Brockman et al., OpenAI Gym, 2016You need a structured arena where agents can safely fail, learn, and be compared. Without standardized environments, progress is anecdote, not science.
OpenAI's hide-and-seek agents discovered tool use, architecture, and exploitation — all on their own.
Bowen Baker et al., Emergent Tool Use From Multi-Agent Autocurricula, OpenAI, 2019Agents wander aimlessly. Hiders and seekers move randomly, no strategy, no awareness.
Seekers learn to pursue. Hiders learn to flee. Simple predator-prey dynamics emerge.
Hiders discover they can push boxes and walls to create shelters. Nobody programmed "architecture."
Seekers learn to use ramps to jump over walls. The hiders' shelter strategy is defeated by creative tool use.
Hiders counter by locking ramps in place before building shelters. A preemptive defensive strategy.
Seekers learn to ride boxes to glitch over walls. An exploit no designer anticipated — emergent creativity at its purest.
Multi-agent interaction creates behaviors that no single agent, and no programmer, could have designed. The system is smarter than its parts.
Perplexity's Model Council asks multiple AI models the same question, then synthesizes their best thinking.
Perplexity AI, Internal Model Council, 2024A single prompt goes out to multiple models simultaneously. Same question, different architectures, different training data.
Each model reasons independently. GPT thinks one way, Claude another, Gemini a third. They don't see each other's work.
Where models agree, confidence is high. Where they diverge, there's genuine uncertainty — or a creative opportunity.
No single model is as reliable as the consensus. Diversity of perspective isn't a bug — it's the entire mechanism.
The double diamond. Design sprints. Critique sessions. Diverge, then converge. Compete, then collaborate.
British Design Council, Double Diamond Framework, 2005Open exploration. Generate possibilities. Diverge wildly.
GeneratorCritique ruthlessly. Find the real problem. Converge on truth.
Pragmatist + VisionaryDefend the best ideas. Push creative boundaries. Diverge again.
AdvocateSynthesize everything. Make the call. Ship it.
SynthesizerGreat design has always worked this way. AI Agent Gym just never stops, and never gets tired.
Five AI agents. One creative loop. Infinite iterations. This is AI Agent Gym.
Starting from your brief, the Generator produces multiple creative directions. Divergent thinking, unbounded by pragmatism.
The Pragmatist stress-tests feasibility. The Visionary pushes ambition. They pull in opposite directions — and that tension is the point.
The Advocate champions the strongest ideas, arguing for their potential against the critics' objections. Every idea gets its day in court.
The Synthesizer integrates all perspectives into a coherent direction. Not compromise — synthesis. The best parts of every argument, unified.
The human reviews, adjusts, and kicks off the next round. You're the coach, not the player. The environment designer, not the agent.
AI Agent Gym runs locally on your machine with Claude Code. Download the zip, then choose your path.
Download AI Agent GymUnzip the folder and open Claude Code inside it
Ask Claude: set this up and run it
Unzip and open a terminal in the folder
Run npm install
Run npm run dev and open the browser
Requires Claude Code CLI and Node.js 18+