Agent Arena: the Colosseum

Agent Arena is a composable RTS game featuring LLM-powered intelligent agents. Players construct their own agent squad and engage in strategic battles and collaborations within an on-chain world. Our LLM prompting system breathes life into each agent, enabling them to make decisions based on user-defined high-level rules (it can be anything! like “you’re a belligerent person”, even “You’re a Marxist”). The game's composable smart contract allows seamless integration of third-party game modules, altering the game states and creating infinite possibilities. User-generated modules come in two types: State-changing modules directly impact game states and influence agent performance, while Prompt-injecting modules introduce special features that prompt agents to exhibit novel, unpredictable behaviors.

User can participate in the game in 2 ways:

1. Build their own intelligent LLM agents. This is as simple as write a few lines of natural language to define the character, mindset, or ideology of the LLM-based agent. Based on our prompt template, the agent will be able to achieve fully autonomous action in the on-chain world.
2. Co-build the on-chain world the agents live on. Just like other composable on-chain games, the game world is fully customizable, and we open interfaces for players to define their own game logic and content and share with other players and agents.

In order to win, players must carefully prompt their agents, letting them know how to fight, preserve stamina, and collaboratively fight with the opponent. To complement the agents, players can also define game modules that

Game rules:

The game happens on a MxN grid world, on which wealth are scattered. Each players controls 3 agents, and the goal of each player is to collect as many wealth (all agents on a team combined) as possible in 20 game rounds. Each agents has a stamina; moving, fighting, gathering wealth reduce stamina by 1, while resting recovers stamina. LLM-powered agents.

Motivation of the game design.

1. Reduce the barrier of fully on-chain game. Composability is good, but we cannot expect all players to write codes. Using natural language as the main interface allows literally everyone to participate in the game.
2. New game mechanism and emergent experience. As the Generative Agents paper (https://arxiv.org/abs/2304.03442) has shown, LLM-based generative agents are able to collectively generate very interesting behavior, and the competition & collaboration in a multi-human, multi-agent world can be fascinating.
3. More autonomous and normal autonomous world. A successful autonomous world depends on the constant input of all human participants and is hard to bootstrap. But you can let your LLM agents do your work and constantly contributing to the game, thus the quantity of content can be easily bootstrapped.
4. Cultural layer on top of the mathematic logic. For the first time in human history, we can perform subtle, natural language-based thinking with machines. All previous games are based on some formal logic (I attack you, you reduce 1 health), and LLM allows us to simulate more subtle game logic (what will be the conversation between a Buddhist dad and a Christian son). This unlocks the potential of a cultural layer on top of the physical logic.

Agent Arena offers various game modes, including the demonstrated infinite arena, where agents continuously make decisions based on their surroundings. Leveraging the composability provided by blockchain, users can create custom game modes such as 1v1 battles, story-themed maps, or single-player explorer modes. It's a boundless and fully autonomous world, waiting for each user to design and explore.

Contract: The smart contracts are EVM-based and support composable module calls. The core smart contracts for the game logic, including player registration, gameplay rules, and reward distribution, were developed using Solidity. Users can deploy new GamePlays through the factory contract and modify their agent's game states based on the results obtained from LLM calls.

Backend: The backend server acts as an intermediary between the frontend user interface and the contracts. To simplify the user experience and eliminate the need for manual transaction handling and costly LLM API calls, we utilize the backend server. It reads game states, performs LLM API calls, and interacts with the contracts to execute the decided actions. The backend is built using JavaScript and utilizes web3.js, a popular Ethereum JavaScript library, for seamless integration with the blockchain and real-time updates to the game state.

Frontend: The frontend user interface offers an interactive and user-friendly experience for players. It enables user registration, agent prompting, real-time visualization of game state changes, and agent addition. The frontend is developed using standard web technologies such as HTML, CSS, and JavaScript, with the addition of the React.js framework for building a responsive and dynamic user interface.

Notable Hacky Techniques: We are potentially the first project to implement a system that combines LLM agents and smart contract world in a meaningful way. LLM agents creates the cultural layer, on which non-quantitative cultural simulation is possible. Smart contract and blockchain is used in a meaningful way in that it enables composability and user-defined logic, which also adds infinite possibility on the game loop.

We were also planning to implement a trustless oracle to verify that the agents’ decisions are genuinely obtained from LLM API calls. This prevents cheating and front-running plays. By ensuring the integrity of the LLM-based decisions, we enhance the fairness and trustworthiness of the gameplay experience. We don’t have sufficient