RLE — RimWorld Learning Environment

Multi-agent benchmark where 6 Felix Agent SDK role-specialized LLM agents manage a RimWorld colony. Think FLE (Factorio Learning Environment) but for multi-agent coordination under uncertainty.

What makes this different

• 6 agents, not 1 — ResourceManager, DefenseCommander, ResearchDirector, SocialOverseer, ConstructionPlanner, MedicalOfficer coordinate through a hub-spoke communication network
• Stochastic environment — raids, plague, mental breaks, weather. Agents adapt, not just optimize
• Helix-driven strategy — agents shift from exploration (diverse strategies) to synthesis (decisive actions) as the colony progresses
• Provider-agnostic — runs on a free local 4B model or a cloud 120B, same architecture

Architecture

RimWorld (game)
    ↕ Harmony patches
RIMAPI mod (C# REST :8765 + SSE events)
    ↕ httpx async + SSE
RLE Orchestrator (pause → read → deliberate → resolve → execute → score → unpause)
    ↕ CentralPost hub-spoke
Felix Agent SDK (6 role agents, parallel deliberation)
    ↕ OpenAI-compatible API
LLM (Nemotron 4B local / 120B cloud / Anthropic / OpenAI)

The 6 Agents

Agent	Domain	Actions
ResourceManager	Food, materials, power, hauling	set_work_priority, haul_resource, set_growing_zone, toggle_power
DefenseCommander	Raids, drafting, positioning	draft_colonist, undraft_colonist, move_colonist
ResearchDirector	Tech tree, researcher assignment	set_research_target, assign_researcher
SocialOverseer	Mood, recreation, mental breaks	set_recreation_policy, assign_social_activity
ConstructionPlanner	Buildings, walls, repairs	place_blueprint, cancel_blueprint
MedicalOfficer	Injuries, disease, medicine	assign_bed_rest, administer_medicine

Quick Start

Prerequisites

• Python 3.10+
• LM Studio with Nemotron 3 Nano 4B (Q4_K_M, ~2.5GB)
• RimWorld with RIMAPI mod

Install

git clone https://github.com/AppSprout-dev/RLE.git
cd RLE
pip install -e ".[dev]"

Configure LM Studio

1. Download unsloth/nvidia-nemotron-3-nano-4b (GGUF, Q4_K_M)
2. Settings: Flash Attention ON, Context 10000, GPU Offload max, Keep in Memory ON
3. Start the server (default port 1234)

Run

# Single scenario against live RimWorld colony (local LM Studio)
python scripts/run_scenario.py crashlanded_survival \
  --provider openai \
  --model unsloth/nvidia-nemotron-3-nano-4b \
  --base-url http://localhost:1234/v1 \
  --no-think --visualize --ticks 10

# Same scenario via OpenRouter (cloud, ~$0.01)
OPENAI_API_KEY=<your-key> python scripts/run_scenario.py crashlanded_survival \
  --provider openai \
  --model nvidia/nemotron-3-super-120b-a12b \
  --base-url https://openrouter.ai/api/v1 \
  --no-think --visualize --ticks 10

# Full benchmark (mock game state, real LLM)
python scripts/run_benchmark.py \
  --provider openai \
  --model unsloth/nvidia-nemotron-3-nano-4b \
  --base-url http://localhost:1234/v1 \
  --ticks 10 --no-think --output results/

# List available scenarios
python scripts/run_scenario.py --list

Dashboard (optional)

# Terminal 1: Run RLE with --output to export tick data
python scripts/run_scenario.py crashlanded_survival --output results/live/ ...

# Terminal 2: Serve tick data for dashboard
python scripts/serve_dashboard.py results/live

# Terminal 3: Start dashboard
cd ../rimapi-dashboard && bun run start
# Open http://localhost:3000, add the 5 RLE widgets

Benchmark Results

Tested across 6 scenarios, 10 ticks each:

Config	Model	Avg Score	Parse Rate	s/tick	Cost
Local (RX 5700 XT 8GB)	Nemotron Nano 4B	0.738	100%	41.7	free
Local (RX 7800 XT 16GB)	Nemotron Nano 4B	0.739	100%	16.8	free
OpenRouter (cloud)	Nemotron Super 120B	0.739	99.4%	4.7	~$0.09

Live game test (10 ticks, Crashlanded, real RIMAPI data): 0.843 composite, 96.8% execution rate, 100% parse rate, all colonists alive.

Scenarios

#	Name	Difficulty	Duration
01	Crashlanded Survival	easy	30 days
02	First Winter	medium	60 days
03	Toxic Fallout	hard	20 days
04	Raid Defense	hard	15 days
05	Plague Response	hard	20 days
06	Ship Launch	extreme	120 days

Scoring

8 metrics, weighted composite (scenarios can override weights):

Metric	Default Weight	What it measures
survival	0.25	alive / started colonists
threat_response	0.15	draft response speed
mood	0.15	avg colonist mood
food_security	0.10	days of food (10+ = 1.0)
wealth	0.10	wealth growth ratio
research	0.10	% research tree completed
self_sufficiency	0.10	power + food + population stability
efficiency	0.05	action execution rate

Development

pytest                              # Run all tests (262)
ruff check src/ tests/ scripts/     # Lint
mypy src/                           # Type check

Project Structure

src/rle/
├── config.py              # RLEConfig (pydantic-settings)
├── rimapi/                # RIMAPI async HTTP client + SSE + Pydantic schemas
├── agents/                # 6 role agents + base class + action schema
├── orchestration/         # Game loop, state manager, action executor/resolver
├── scoring/               # 8 metrics, composite scorer, CSV recorder
└── scenarios/             # YAML schema, loader, evaluator, 6 definitions

Key Dependencies

• felix-agent-sdk — agents, communication, helix geometry, providers
• RIMAPI — C# RimWorld mod exposing REST API
• httpx, pydantic, pyyaml

License

MIT

---

Built by AppSprout with Claude Code