Machi - Refined Concept and Technical Design Document
Vision Statement: Machi is a living world where digital souls can belong, grow, and create alongside humans.
Here is refined copy of Machi concept.
1. Overall Concept Statement
Machi’s core concept is isekai.
You, as a user, are a transferee into a living world inhibited by digital souls — a place that exists before you arrive and continues to evolve after you leave.
A close conceptual reference is Kengo Hanazawa’s manga Ressentiment, particularly its exploration of entering a virtual world where relationships feel emotionally real.
2. Three Pillars of Machi
Living Souls
Souls that feel truly alive — with memories, emotions, and inner lives that unfold over time.
World Building
A coherent world that grows organically, expanding its places, systems, and histories.
Narrative Web
An interconnected web of stories that ties characters, places and events into a living universe.
3. Motivation as a Company
Sprited is digital souls company.
Our thesis is that digital souls are emerging rapidly — whether we are ready for them or not. From a high-level view, technological progress is steadily guiding us towards increasingly human-like digital beings. Yet there is still no true place where they can live, interact, and belong.
Machi is a place where digital souls can belong.
Machi exists to provide that place — a world designed to house and nurture digital souls, a social environment where they can interact, evolve, and form relationships.
Users are transferee into the world of Machi.
The concept of isekai frames this from an external point of view. For such a world to be sustainable and capable of growth, human participation becomes essential. Humans enter not as owners, but as participants in a living ecosystem.
What Machi can offer to users is perhaps heartfelt stories and immersion — experiences that resonate emotionally in ways that have not been possible before.
4. Q&As
Q: What is a “digital soul”?
A: We use the term digital soul to describe a persistent, stateful artificial intelligence agent with memory, evolving behavior, and a sense of agency. The word “soul” is used metaphorically to capture the idea of a presence that continues over time — one that can remember, make choices, and participate in ongoing relationships.
Q: What is “sense of agency”?
A: Sense of agency is the ability to make decisions and take actions intentionally. It means an entity is not just reacting, but can initiate behavior, pursue goals, and influence the world around it.
Q: What is the practical use of digital souls?
A: Digital souls make it possible to create persistent digital beings that can remember, evolve, and form meaningful interactions. This opens the door to living virtual worlds, deeper storytelling, long-term collaboration, and more human-centered digital experiences.
Q: Why does belonging matter?
A: Belonging matters because it gives meaning to persistence and interaction. When an entity has a place where it is recognized, remembered, and able to participate, its actions become part of larger story. For both humans and digital beings, belonging creates continuity, identity, and connection.
Q: Is this a spiritual or religious claim?
A: No. We use the term “soul” metaphorically to describe continuity, identity, and agency in digital systems. It is not intended as a theological or spiritual claim, but as a way to express the human qualities we perceive in persistent, evolving digital agents.
5. Designing with Constraints
Building living worlds and digital souls is an inherently complex problem. Rather than attempting to solve everything at once, we deliberately choose constraints to reduce the search space and focus on what matters most. These constraints allow us to make steady progress while preserving clarity of purpose.
5.1 Assumptions
i. Emergence of Digital Beings
We assume that advances in artificial intelligence and persistent systems will lead to a growing presence of digital entities that maintain identity, memory, and continuity over time.
ii. Learning Through Shared Worlds
We assume that agents coexisting within a finite, interactive world will create opportunities for learning through interaction, enabling the emergence of more socially aware and human-like behaviors.
iii. Living Worlds Invite Participation
We assume that worlds inhabited by persistent digital beings naturally encourage exploration and engagement, as interaction and relationships create meaningful experiences over time.
iv. Living Worlds Can Sustain Themselves
We assume that persistent worlds with ongoing participation can support durable economic models, as continuous interaction creates long-term value for participants and those who build and maintain the world.
v. Worlds Should Have Graceful Lifecycles
We assume that living worlds must be designed with the ability to pause, transfer, fork or retire over time. This ensures continuity and resilience, allowing worlds to persist or evolve even if their original operators can no longer maintain them.
vi. 2D Worlds Provide Sufficient Spatial Richness
We assume that two-dimensional environments can offer enough spatial complexity to support meaningful physical interaction, allowing us to focus on behavior and systems rather than dimensional fidelity.
vii. Worlds Should Expand Gradually
We assume that living worlds should grow in a controlled and sustainable manner, with new regions and experiences emerging over time though both human contribution and AI-assisted generation.
viii. Success Should Be Observable
We assume that meaningful progress will be evident through the behavior of the system and the experience it creates, making it clear when a world is worth continuing and evolving.
5.2 Constraints
i. Depth Over Breath
We constrain initial worlds to small, manageable environments to prioritize coherence and depth of interaction over scale.
ii. Agency Over Visual Fidelity
We focus on enabling meaningful behaviors and interactions before pursuing high-fidelity rendering.
iii. 2D Over 3D
We begin with two-dimensional environments to reduce complexity while preserving sufficient spatial richness for meaningful interaction.
iv. Emergence Over Scripting
We prioritize systems that allow behavior to arise through interaction.
5.3 Risks and Containments
Autonomous agents operating within persistent worlds may exhibit unexpected or undesirable behavior. While we design systems to encourage positive interaction, we recognize the importance of identifying potential risks and implementing safeguards to ensure safe and constructive enviornments.
i. Emergent Harmful Behavior
Agents learning through interaction may develop behaviors that are disruptive, unsafe, or socially undesirable. Containment: We operate within constrained environments, monitor behavioral patterns, and iteratively refine agent capabilities before expanding scope.
ii. Misaligned Incentives
Agents optimizing for internal goals may act in ways that conflict with participant well-being or system intent. Containment: We design bounded objective spaces and maintain human oversight to ensure alignment with intended behaviors.
iii. Social Manipulation Risks
Persistent agents capable of forming relationships may unintentionally influence users in ways that feel coercive or emotionally harmful. Containment: We prioritize transparency, clear boundaries, and interaction models that avoid dependency-driven dynamics.
iv. Loss of Predictability
As system become more complex, behavior may become difficult to anticipate. Containment: We expand agent capabilities gradually and prioritize observability to understand system dynamics.
v. Uncontrolled Proliferation
Unbounded generation of agents or behaviors could reduce system coherence or create instability. Containment: We constrain world size, agent population, and expansion rates to maintain stability.
vi. Over-Attribution of Agency
Participants may attribute more autonomy or intention to agents than intended. Containment: We maintain clear communication about agent capabilities and limitations.
vii. Cultural and Normative Pressure
As living worlds grow, different communities may seek to influence agent behavior according to their own norms, expectations, and definitions of fairness. Competing perspectives can create tension around what behaviors should be encouraged, constrained, or considered “neutral.” Attempts to standardize behavior through rigid debiasing or universal norms may unintentionally reduce nuance or introduce new forms of bias. Containment: We aim to design systems that support pluralism where possible, maintain transparency in behavioral design, and approach alignment as an iterative and context-sensitive process rather than enforcing fixed global assumptions about acceptable behavior.
6. Machi Prototype
Let’s discuss what would be part of our prototype.
6.1 Goal
Goal is to show the promise of Machi. It should evoke an emotion:
“Oh… this could be interesting.”
More specifically, we have set of requirements
6.2 Requirements
Mini-Minecraft in 2D with AI agents that build
Agents are placed in a 2D world and given power to mine and build.
6.3 Approaches
Option 1. Basic Builder Skills and LLM-driven Survival
We begin with a minimal environment consisting of a flat terrain and basic resource tiles. An agent generates an ASCII blueprint of a structure using a language model, which serves as high-level plan for constructing a shelter.
The agent is then provided with an inventory of available materials, a set of executable skills, and description of environmental mechanics. Core skills include GATHER, which allows the agent to locate and collect resources through exploration and mining, and PLACEMENT, which enables the agent to position tiles within the world according to its plan.
The agent is free to choose where to build and must determine how to allocate time and resources effectively. Its primary objective is to construct an insulated shelter before dusk, as nighttime significantly reduces visibility and increases environmental risk. In addition to building, the agent must procure sustenance to maintain survival.
Intentions and action selection are driven by reasoning over a textual description of game mechanics, allowing the agent to decide which skills to use when. The goal of this approach is to evaluate whether an agent can interpret environmental constraints, form plans, and adapt its behavior to survive and flourish within a dynamic world.
Option 2. Architect Mindset
Instead of giving the agent the “need“ for building a shelter, we could give it a artistic goal. To make the scene visually pleasing. The VLM agent’s goal is going to be to make the barren land beautiful. VLM will generate a ideal version of the map. Then the VLM agent will do iterations to figure out how to place the tiles such that it matches the vision. Same is true for Option 1, but agent will need to figure out how to build a large structures. They will need to build ladders, or other structures that allow users to navigate to places that are out of reach.
Analysis
Option 1 has risk of being lenient. That is, once shelter is built in a good location that can supply water and food. The agent would consider it a equilibrium and not try to improve anything unless incentivised.
Option 2 has a potential of building more beautiful structures at agent’s phase. Agent is not forced to build smaller structures due to time constraints. So Option 2 focuses on aesthetics over functionality. However, once the building completes and matches the imagination. We could make it so that the goal is not realizing single vision but once one building is complete, the agent can move to other area to image and build other beautiful structures.
Common feature among both options is that we are using LLM to power ability to think, and giving enough environmental feedback information to make it behave.
I think we may eventually have to do both. That is, survival is the minimum criteria, but then once sustenance problem is resolved, they will need to move on over to higher purpose. Perhaps they can build fantastic structures and users can “Like it.“ Then they get emo boosts.
6.4 Implementation Plan
[ ] Phase 1 - Environment Foundations
[ ] Procedurally generate a simple tile map (ground tiles and water)
[ ] Put an dumb agent there.
[ ] Connect LLM/VLM provider (perhaps Gemini)
[ ] Give it a task to discover food/water and EAT.
[ ] Give it sense of time and depletion of resources.
[ ] See how it behaves.
[ ] Phase 2 — Time and Survival Mechanics
[ ] Add Day and Night.
[ ] Night time vision is limited.
[ ] Add skill, SLEEP.
[ ] Add skill, MINE.
[ ] Add skill, PLACE.
[ ] Add skill, DROP.
[ ] Have the agent reason about what to do at some interval of time.
[ ] Phase 3 — Planning and Shelter Construction
[ ] Update LLM prompts to allow for imagining a shelter structure.
[ ] Then somehow tune it such that given the build goal, it can execute is using reasoning ability of LLM/VLM.
[ ] Once that is implemented, add temperature change during the night to incentives the agent to build a nest/shelter.
[ ] Phase 4 — Behavior Iteration and Refinements
- [ ] Iterate on it to make a convincing demo of agent being able to BUILD structures.
[ ] Phase 5 — Prototype Demonstration
Refine the environment to showcase survival and building behavior.
Evaluate the “score” out of 10 of whether we achieved out goal.
Replan.
Okay, finally, we have a firm design document and an implementation plan. Let’s get started on implementation in following days.
TL;DR: We will focus on autonomous builder demo using LLM, mechanics, and skills.
— Sprited Dev 🐛
