Create sandboxed on-chain simulation environments where agents learn and evolve strategies in controlled, game-theoretic scenarios (e.g., resource allocation, prediction markets, collaborative problem solving).

Key Components:

• Protocols for running agent “games” directly in smart contracts.
• Large amounts of agents should be deployable per simulation (1k to 100k agents per simulation)
• On-chain, reward-driven environments for training reinforcement learning agents.
• Tooling and APIs to allow AI developers to quickly spin up simulations and scenarios.
• Ambition: A rich, on-chain testbed that produces verifiably trained agents, whose learned policies can then be deployed confidently into production DeFi, NFT, or real-world data scenarios.

Practical approaches

1. Realms - Eternum is an onchain civV-like game where players manage resources and play to grow their onchain civilisation. This is the perfect framework for making a competitive onchain simulation for agents. Realms has its own defi system and a complex model of resources and actions which can be replicated and explored easily.
2. https://www.influenceth.io/ is a similar game where players must create teams and assign them to extract resources and explore space.
3. Prediction markets—Replay prediction markets with the Wayback Machine to analyze information in real-time. Partner with X to test this. (We need a feed of social media data that we can re-run.) Implement info-trading backtesting in the smart contract.
4. HummingBot wrapper - similar to the prediction market backtest, set up a TEE-based humming bot backtest framework where agents can arbitrarily configure strategies for over 100s or 1000s of iterations.