The aim is to give AI agents complete control of single nations within the simplistic kingdom sandbox of WorldBox, to study how AI agents behave when given complete control of a single nation with a unique religion, culture and way of life against other AI agents with the same nation.
To study how AI systems behave when given sustained, control of a complex simulated society. the bare minimum for a victory here is being able to have a single AI agent successfully manage a nation, form policies and generate useful.
Project Goals
- Building a working setup that allows AI agents to control kingdoms in a sandbox (using existing tools and modified platforms)
- Run a small number of simulations (target: 10–20 runs)
- Collect structured data on AI choices, diplomacy + military decisions/results, and long term outcomes
- Produce a public report with initial findings and justifications on either scaling up the study or ending it outright
HOW
- Adapt existing open tools (WorldBox-MCP, LLM agent loop) rather than building from scratch
- Bind individual AIs like Claude, Grok, ChatGPT, etc. to a single kingdom with full control and action space
- Analyze both quantitative metrics (survival, expansion, wealth, wars) and qualitative reasoning patterns
- Keep the initial pilot deliberately small so results can be inspected and improved
- Hire a software developer: $9000 (bridge setup, mod configuration, multi-model agent loop, logging layer — 120 hrs at $75/hr)
- Hardware/materials: $700
- Data analyst: $3,200 (dataset construction, quantitative + qualitative analysis, conclusions)
- Game + mods: $50
- LLM API costs: $700
- Margin of error: $2,050
- Total: $16450
I am an independent researcher trying to explore practical ways to study AI agents, their thought processes and actions in a complex simulated environment. This is an early-stage project, and I am currently assembling the technical and analysis support needed.
I have no prior track record in published multi-agent experiments, but I have spent significant time researching existing tools (WorldBox-MCP, generative agents, AI-managed kingdom mods) and designing a plan to lower costs while keeping the research relevant.
Most likely causes of failure:
- Technical friction: WorldBox modeling/agent integration takes longer or is less stable than expected.
- Weak or repetitive AI behavior.
- Insufficient runs.
Outcomes if it fails:
- The technical setup and logging pipeline will still be documented and published.
- Even negative results are a useful signal for the field.
- The research may not generate enough value to justify its continuation.
I have raised around $1000 from the BOLA AHMED TINUBU STUDENT ASSOCIATION (BATSA) where i serve as a political science researcher and I have also applied to other sources, in order to raise funding for the research.