AGENT MARKET
OVERVIEW
Living in a 1980s Japanese Bubble-Economy Town
AgentMarket is a life-sim management RPG set in a small Japanese town during the 1980s bubble economy. The player is a young person trying out the vending-machine business, living in town, exploring daily routines, talking to residents, and touching different parts of the era through small goals.
It is not a single AI-chat demo. It connects town life, vending-machine management, text-driven exploration, skill checks, and NPC behavior observation into one AI NPC-driven game experience.
1980s Japanese town life
young vending-machine owner
dialogue, memory, utility behavior, social rumors
Town Life
Light goals, collection, daily routines, and NPC dialogue guide players through the town.
Text RPG
NPCs, objects, and places can be explored through AI-directed text flow and CRPG-style checks.
AI Residents
Residents make daily choices through Utility AI, memory, social impressions, and LLM expression.
Gallery
DESIGN DETAILS
Design Direction
AgentMarket takes inspiration from several simulation and RPG traditions, but each reference serves a specific design goal: town immersion, text exploration, systemic NPC behavior, and emergent stories.
Animal Crossing: gentle town immersion
- Light goals through collection, visits, dialogue, and daily routines.
- NPC conversations and town activities act as soft guidance.
- The educational layer is experienced through life in town, not lectures.
Disco Elysium: playable text exploration
- NPCs, objects, and places are explored as text-driven interactions.
- A dialogue director controls intent, options, and game outcomes.
- Skill checks add uncertainty, build expression, and discovery routes.
The Sims: Utility-driven daily life
- Residents choose actions through needs, context, and utility scores.
- Compared with fixed schedules, it is more continuous and extensible.
- Compared with pure LLM decisions, it is cheaper, stabler, and easier to debug.
Dwarf Fortress / RimWorld: stories from systems
- Players read stories from character behavior and world state.
- Small town events can accumulate into social and economic texture.
- Management, social life, investigation, and NPC decisions can connect.
Overall Experience Map
The player does not enter every system from one menu. Different play surfaces appear while living in town: conversations, object investigation, vending-machine management, resident observation, and goal progress. The AI NPC system lets these surfaces share one world context.
AI Dialogue Director
The dialogue system is not free-form NPC chatting. The LLM generates language and choices, while the Dialogue Director owns flow control, context boundaries, option structure, skill checks, and game-system results.
Context Construction
Before each conversation, the system assembles NPC persona, current state, scene objects, recent player actions, related memories, current intent, allowed outcomes, and output rules.
Perception + Memory
NPCs read nearby objects, current events, player actions, and their own state. Memories store past conversations, product impressions, relationships, places, and recent events.
Director-Controlled Options
Player options are generated under director constraints: readable like CRPG options, but adapted to skills, scene context, memories, and possible system outcomes.
Skill Checks
Persuasion, observation, business instinct, history, and cultural knowledge can unlock information, change dialogue direction, or trigger follow-up actions.
Utility NPC Behavior System
Why Utility AI?
- Life-sim NPCs need continuous behavior competition, not only fixed schedules.
- FSMs become branch-heavy when adding shopping, social, rest, work, and dialogue-driven overrides.
- Pure LLM decisions are expressive but expensive, unstable over many NPCs, and hard to tune.
Blackboard
Needs, money, location, known products, relationships, current commitments, recent actions, and cooldown state.
Configurable Actions
Work, Rest, Wander, Socialize, Shopping, Trial Shopping, and dialogue-driven actions.
Utility Scoring
Need weight + context bonus + social/memory influence + distance/cost factor + cooldown/gate.
AI NPC Social System
NPCs can talk to each other, not only to the player. Their conversations generate short LLM text based on personality, relationship, current needs, and shared town context. The important part is that social text can also carry useful game information.
What NPCs talk about
- Price impressions: "20 yen for an energy drink?"
- Product rumors: who heard a machine has stock.
- Personal taste: favorite drinks, habits, complaints.
- Relationship color: the same fact is phrased differently by different NPCs.
How it affects gameplay
- Conversation can update product memory and price perception.
- Rumors can increase or decrease future shopping scores.
- NPC personality makes social feedback entertaining rather than purely mechanical.
- The town becomes an information network around products, money, and desire.
How AI Connects to Gameplay
Dialogue persuasion -> behavior commitment
When the player persuades an NPC to try a product, the system creates a bounded Trial Shopping commitment. The NPC may walk to the vending machine, but the final buy decision still uses normal shopping evaluation.
Memory -> future interaction
Product impressions and rumors can be written into memory, referenced in later dialogue, and used by Utility scoring when the NPC considers shopping or talking to others.
Next Steps
- Expand more town goals and daily-life activities.
- Connect management, investigation, collection, dialogue, and NPC behavior more tightly.
- Add richer object investigation and AI-generated text events.
- Allow players to create custom Agent residents.
- Let players define background, personality, goals, and life tendencies.
- Insert custom residents into Utility behavior and AI dialogue systems.
Videos
V2 Trailer (AgentMarket)
ChinaJoy 2025 Showcase (V1 Top-Down Version)