AI Villagers -- A game to convince, compete for and collaborate with your AI NPC

Hyperion HyperHack Ideathon
1

Project Name

AI Villagers / BotNextDoor

Problem Statement

Most AI-powered games today use artificial intelligence as a background tool—automating movement or generating filler content —rather than making it central to gameplay. Players rarely engage in meaningful, emotionally resonant interactions with AI characters who possess memory, autonomy, and preferences.

AI Villagers is here to fill the gap. We’re building a new kind of multiplayer strategy game with dynamics that is not previously possible before the age of LLMs. In this world, you must persuade intelligent, opinionated AI villagers to join your community—not through combat or quest completion, but by building relationships and negotiating values.

Solution Overview

AI Villagers is a persistent online multiplayer game that uses AI NPCs as central agents, not background tools. These NPCs make independent decisions based on player actions, conversation tone, and their own internal personalities. Every AI decision is stored on-chain for transparency and verifiability, leveraging Metis SDK and Alith for dialogue inference and logic execution.

Players must strategically build environments, manage relationships, and hold intelligent conversations to convince AI NPCs to live—and stay—in their village. With AI-powered decision-making, personality modeling, and persistent conversational memory, the game becomes a social simulation of diplomacy, resource management, and emotional intelligence.

Core Gameplay Mechanics

1. Village Building (Strategic Design Layer)

Players build out their village using a grid-based system. Each tile contributes to environmental stats such as comfort, pollution, efficiency, and aesthetics. Different villagers are drawn to different traits (e.g., minimalists vs. techies vs. nature lovers). There’s no one-size-fits-all solution—your environment must align with your target NPCs.

Some NPCs also carry inventories containing unique resources or “building blocks” that only become available to the village they reside in. These inventories refill at a fixed rate, adding strategic value to recruiting and retaining specific villagers.

2. NPC Decision-Making (AI-Powered Logic + On-Chain Scoring)

To prevent players from gaming AI decisions, the recruitment process relies on a hybrid scoring system:

  • Base Score (Non-AI, Deterministic): Based on static environmental factors like resource availability, pollution levels, or infrastructure.
  • NPC Affinity Score (AI, Context-Free): Determined by the tone and sentiment of player conversations—players who are rude or manipulative will get penalized.
  • NPC Preference Score (AI, Context-Aware): Based on deeper personality modeling and evolving preferences over time.

These scores are combined and compared against a threshold to determine whether the NPC joins the village. Final decisions and scores are made via on-chain AI inference using Alith, ensuring transparency and resistance to manipulation.

To interact, players must choose:

  • [Chat]: Send a message to the NPC (on-chain, costs a small ETH fee per message).
  • [Convince]: Trigger an on-chain decision using prior interactions and environmental data.

There are daily chat limits, encouraging thoughtful and strategic engagement over spamming.

3. Strategic Recruitment & Competitive Diplomacy (PvP + PvE Blend)

Villagers can only reside in a limited number of villages simultaneously. If a new player convinces an NPC already housed in another village, the system will drop the village with the lowest cumulative affinity score—causing migrations that simulate soft PvP dynamics. NPCs can migrate over time if their satisfaction falls, adding tension and competition.

Some of the features below still need to be fleshed out, but ideally players can:

  • Form guilds and coordinate recruitment efforts (TBD)
  • Trade resources freely with others to build alliances.
  • Use NPC relationships (stored as a dynamic matrix of inter-NPC affinity scores) to attract clusters of villagers who like one another.

Villages may also be geographically mapped on a larger grid. We’re exploring how proximity affects competition and migration threshold.

4. Conversational Strategy & Social Risk

A key innovation is the publicly queryable conversation history stored on-chain. Players can:

  • Reference past conversations from themselves or others.
  • Use excerpts as “evidence” during persuasion attempts.
  • Your word may be used against you when others try to convince their NPC to migrate

This mechanic encourages players to treat NPCs with respect and emotional intelligence, as every word may come back to affect gameplay. We want to make a game such that talking to an AI in this game is more like talking to a real person you would meet than issuing commands.

Leaderboard & Rewards

Players are ranked by the weighted cumulative total stay-time of NPCs in their village. The longer an NPC stays, the more points you earn. Top-ranked players may receive NFTs, token rewards, or cash prize.

This reward system encourages not only recruitment but long-term relationship building—a village that NPCs want to stay in.

Tech Stack

  • Frontend: Next.js
  • Game Engine: Phaser
  • Backend: Django / Express
  • On-Chain AI: Alith + Metis SDK
  • Indexing: The Graph

Community Engagement

Players and early community members will be invited to beta test and submit suggestions for NPC personalities, character traits, and village archetypes. Community-designed characters may be added to the official pool of villagers.

10 Likes
2

Amazing idea bro, This will really add a new layer of excitement as all the interactions will be different everytime we change something

These NPCs make independent decisions based on player actions, conversation tone, and their own internal personalities

  • If every NPC makes an independent decision won’t it be hard because every player will have some other reaction to some event happening or will it be like a choice based?
  • How will the belief or the behaviour of the villagers change over time as we progress in the game

Every AI decision is stored on-chain for transparency and verifiability, leveraging Metis SDK and Alith for dialogue inference and logic execution.

  • How do you ensure on-chain AI decision-making remains deterministic, given that models like Alith and LLM-based inferences are inherently probabilistic?

Sorry if i asked too many questions :victory_hand:

6 Likes
3

btw I forgot to mention but here are some content you can check out :victory_hand:

General Documentation:

Some Scalability Content For Scaling the game:

  • Why Most Web3 Games Fail at Scale (and How To Fix It) | Link
  • What Fully On-Chain Games Mean for Gamers & Developers | YT Video Link
  • Lessons in Hyperscaling Onchain games - Web2 meets Web3 - Yijia Chen | YT Video Link

Blogs:

Some good startups based on On-chain Gaming

  • Mythical Games
  • Atmos Labs
  • Send Arcade by Solana
5 Likes
4

Thank you for all the questions! Those are great question!

  • If every NPC makes an independent decision won’t it be hard because every player will have some other reaction to some event happening or will it be like a choice based?
  • How will the belief or the behaviour of the villagers change over time as we progress in the game
  • Most of the NPC’s decision is based on their personality/environmental preference, their inventory vs the “environment/village” that the player created and how they choose to approach the NPC to convince them.
  • As for events, event will cause state change environment and inventory(eg a npc gift something to another npc ), so it will indirectly influence the decision. But we probably won’t include a memeory of all event that happened during AI inference.
  • The relationship amongst the villager/player may change (as in affinity score may increase or decrease). But for simplicity their base “personality/preference” attributes will probably remain constant. But a changing belief is indeed something we can work on when after we develop a working MVP.

How do you ensure on-chain AI decision-making remains deterministic, given that models like Alith and LLM-based inferences are inherently probabilistic?

Ideally we want the AI decision itself be probabilistic, that will make the game more fun. The deterministic score(base score) I mentioned will be a non-AI algorithm.

(btw I wouldn’t say LLM-based inferences are inherently probabilistic. In reality, if we use our own model with fixed model state, inputs, and parameters , temperature=0, mathematically speaking a LLM is just a large neural network and will produce deterministic outputs after doing a bunch of matrix multiplication with the input. But that’s a bit of a sidetrack here. )

How will the product scale when there will be 10000 people playing and over 100 villagers each having different reactions to certain events and all the data will be stored on-chain

We will probably issue village as NFT and cap at at for example 1k to ensure we have a manageable amount of players.

I don’t think there would be a problem with storing all those data on chain, as separate transactions. We simply won’t pass all those data to AI during inference state, the villagers won’t “remember” everything. But we can allow players to [select] certain conversation snippet as parameter during their [convince] action and that will be part of the strategy.

5 Likes
5

Thank you for all the info, I’ll definitely check those out!

Some of those articles looks like very interesting reads :heart:

5 Likes
6

glad to be of some help :victory_hand:

5 Likes
7

I love this idea mate, and I love the gaming ideas in this forum, AI Villagers is another one that I would really love to play.

8 Likes
8
  1. Can players influence NPC evolution over time, or are personalities fixed?
  2. What protections exist to prevent Sybil attacks or multi-wallet farming of NPCs?
7 Likes
9

Glad to hear that :heart: Hope I can build out a solid game within the given timeframe!

I was excited to see in the doc on Hyperion high performance network+ Alith AI combo that essentially allows for on-chain AI inference, so I want to make game mechanics that wasn’t previously possible.

I think introducing AI as NPC and provide them with essentially the “voting power” , it could really help prevent a “richest player/big influencer wins everything when it comes to voting” scenario, which is common in web3, and give everyone a fair chance to compete. (My initial idea is to have a “Run for mayor” and NPC vote for you type of game but I the current setup is more suitable for gamification. )

4 Likes
10

Loved the idea !
I have two questions regarding the proposal.

  1. Given that conversational history is stored on-chain and can be publicly queried and used as “evidence” in persuasion attempts, what specific mechanisms or features will be implemented to prevent or mitigate potential negative meta-game behaviors, such as coordinated harassment campaigns against specific NPCs, or players “ganging up” on others by strategically using their past conversations out of context?

  2. The “NPC Preference Score” is described as being based on “deeper personality modeling and evolving preferences over time.” How will the game ensure the long-term diversity and unpredictability of these evolving preferences for a large number of NPCs, preventing them from becoming stale or converging into predictable patterns that players can easily exploit, even with the on-chain AI inference?

Looking forward to see it implemented. all the best

Thank you

5 Likes
11

Hey priyankg! Those are great questions!

  1. Can players influence NPC evolution over time, or are personalities fixed?

The personalities are assumed to be fixed for now for simplicity. (But if we manage to build a working MVP early this may be subject to change).

The relationship between NPCs could evolve over time though and the NPC’s inventory can be affected by events in game.

  1. What protections exist to prevent Sybil attacks or multi-wallet farming of NPCs?

I think villages will be sold as NFTs and having a price on that will prevent spam to some extent.

Players are allowed to own multiple villages and share resource to each other. But ultimately sybiling would not give them that much of the advantage since Players are ranked by the weighted cumulative total stay-time of NPCs in their village.

So even if they own multiple villages they will need to essentially compete for NPC. This will just increase their cost if they own too many villages.

5 Likes
12

Thank you for taking the time to read through my proposal and asking such thoughtful question!

I’ll definitely keep those in mind as I move into the building phase.

  1. Given that conversational history is stored on-chain and can be publicly queried and used as “evidence” in persuasion attempts, what specific mechanisms or features will be implemented to prevent or mitigate potential negative meta-game behaviors, such as coordinated harassment campaigns against specific NPCs, or players “ganging up” on others by strategically using their past conversations out of context?

I think slandering and taking words out of context and use it as an advantage is part of human nature and this can be interesting part of the game as well!

But you did brought up a good point if people are harassing NPC at large scale it would bring too much negativity to the game. Part of the score (which determines whether npc accepts the invite) is determined by the sentiment analysis of the conversation, and a highly negative conversation can result in a negative score in this part, and I hope this can prevent people from bringing up too much negativity in the conversation. We can probably tune the weighting of this portion based on how the user plays the game.

I’ll take this into consideration during building.

  1. The “NPC Preference Score” is described as being based on “deeper personality modeling and evolving preferences over time.” How will the game ensure the long-term diversity and unpredictability of these evolving preferences for a large number of NPCs, preventing them from becoming stale or converging into predictable patterns that players can easily exploit, even with the on-chain AI inference?

This is a great question as well!

I think one "easy " solution is that we can periodically retire some of NPC (probably the least popular and some of the most popular ones) and introduce brand new NPC or brand new resource with different initialization & overall distribution periodically. So player will have to constantly modify their village based on the changing NPC dynamic.

We could also compensate player with some base score if we remove an NPC residing in their village.

3 Likes
13

i think i found a GEM here .. hehehe

3 Likes
14
  • How will you ensure that the on-chain costs for player-NPC interactions don’t discourage engagement or create pay-to-win dynamics?
  • What measures are in place to prevent players from exploiting the affinity scoring system to unfairly recruit or block NPCs?
  • How do you plan to scale the on-chain AI computations as the player base and NPC interactions grow?
2 Likes
15
  • How will you ensure that the on-chain costs for player-NPC interactions don’t discourage engagement or create pay-to-win dynamics?

Yep the tradeoff of making a fully/mostly on chain game is that user will have to pay gas for their actions. But I think on L2 the gas is generally low enough.

To mitigate the issue I think we will cap the number of conversation and action the players can make each day, so that strategy and quality of their action > quantity.

  • What measures are in place to prevent players from exploiting the affinity scoring system to unfairly recruit or block NPCs?

Could you elaborate a bit on what would be considered as “unfairly recruit or block NPCs”?

  • How do you plan to scale the on-chain AI computations as the player base and NPC interactions grow?

Great question! Basically we can ensure the complexity will be O(1) instead of O(N) if we don’t pass in all the previous conversation as context.

What makes this scalable is that past conversations indirectly influence gameplay through state changes of inventory and environment, so we kind of evaded the issue by having to pass all history to AI while making sure past action does have impact on future.

So the NPC decision-making process relies on:

  • The latest player message (O(1))
  • Selected snippet of past dialogue, with a cap of how many you can pass in (O(1))
  • And the current village/inventory state, which has already absorbed the consequences of previous conversation/action.
4 Likes
16

Thank you for the thorough answers and the clever approach to keeping AI computation scalable!

Just a quick question: How do you plan to handle potential abuse from players who might try to game the daily conversation/action limits?

2 Likes
17

Good question!

The conversation/action limit is a hard limit that will be enforced on-chain. Each “action” that can cause state change in game will be an transaction.

But you did remind me of something, it would probably be good to enforce a max character/message limit to ensure people don’t pack too much content into each conversation.

3 Likes
18

Makes sense! Enforcing both action and message limits sounds like a solid way to keep gameplay fair and on-chain costs predictable.

2 Likes