ihere04u/business-strategy-env

🏢 Business Strategy Simulation Environment

An OpenEnv-compliant real-world environment where an AI agent acts as CEO, making quarterly strategic decisions to grow a company. Features stochastic market dynamics and multi-objective reward optimization.

Built for the OpenEnv Hackathon — Round 1 · Live API Docs

🌍 Why This Environment?

Business strategy is a genuine real-world task — companies live and die by quarterly decisions on hiring, marketing, pricing, and R&D. This environment models those decisions with:

• —Stochastic market dynamics — random noise simulates real market unpredictability
• —Interdependent state variables — actions have cascading effects (e.g. cutting costs reduces quality, reducing satisfaction, reducing market share)
• —Multi-objective trade-offs — agents must balance short-term profit vs long-term growth
• —Partial progress rewards — dense reward signal every quarter, not just at episode end

🎯 Tasks

🔧 Action Space

10 strategic actions, each with an optional amount parameter (default: $5,000):

👁️ Observation Space

This environment returns both raw business metrics and higher-level strategy signals.

{
  "revenue": 50000.0,
  "costs": 35000.0,
  "profit": 15000.0,
  "market_share": 0.10,
  "employees": 20,
  "customer_satisfaction": 0.70,
  "marketing_budget": 5000.0,
  "rd_investment": 2000.0,
  "product_quality": 0.65,
  "profit_margin": 0.30,
  "cost_efficiency": 0.30,
  "growth_signal": 0.40,
  "profit_trend": 0.0,
  "last_reward": 0.0,
  "risk_level": 0.70,
  "strategic_health": 0.37,
  "growth_momentum": 0.07,
  "decision_quality": "neutral",
  "quarter": 1,
  "max_quarters": 4,
  "done": false,
  "reward": 0.0,
  "message": "Q1: Profit=$15,000 | Market=10.0%"
}

📡 API Endpoints

🏆 Reward Design

All rewards are partial progress signals — no sparse binary end-of-episode rewards:

• —survive: profitable_quarters / total_quarters — scores every quarter
• —grow_market_share: min(final_share / 0.20, 1.0) + early completion bonus
• —scale_profitably: 0.6 × revenue_score + 0.4 × satisfaction_score — weighted multi-objective

Undesirable behaviors are penalized:

• —Bankruptcy (profit < -$50,000) → early termination
• —Over-hiring with no revenue → costs spiral punishes poor decisions
• —Cutting costs repeatedly → product quality degrades, reducing future revenue

🧠 Advanced Learning Dynamics

This environment is intentionally designed to challenge decision-making agents through:

1. Multi-Objective Optimization

Agents must balance:

• —Profitability
• —Market share growth
• —Customer satisfaction
• —Cost efficiency

2. Delayed Rewards

Investments in R&D improve future revenue rather than immediate outcomes.

3. Strategic Trade-offs

Each action has both positive and negative consequences:

• —Expanding markets increases growth but reduces satisfaction
• —Cost cutting improves margins but degrades product quality

4. Stochastic Environment

• —Economic cycles affect revenue unpredictably
• —Competitor pressure reduces market share dynamically

5. Non-Linear Reward Shaping

Rewards include:

• —Trend-based bonuses
• —Strategic diversity incentives
• —Penalties for repetitive or short-sighted decisions

6. Failure Cascades

Poor decisions (e.g., low satisfaction) trigger compounding negative effects.

7. Decision Feedback

The environment exposes decision quality signals such as decision_quality and a final_summary at episode end.

This creates a realistic environment requiring long-term planning, adaptation, and strategic reasoning.

📊 Baseline Scores

Scores from the included rule-based baseline agent (baseline.py):

Note: Baseline performance is intentionally stochastic and may vary across seeds.

🤖 Agent Strategy

The included rule-based inference agent (inference.py) uses adaptive task-specific logic:

• —Survive: Maintains profitability by cutting costs when needed, then rotates between growth actions
• —Grow Market Share: Progressively increases market reach—expands when low, invests in marketing, then launches products
• —Scale Profitably: Balances quality, satisfaction, and growth—invests heavily in R&D, then scales with pricing and expansion

The agent includes anti-repetition safety to avoid over-using the same action, ensuring diverse strategy execution. The agent is fully rule-based and does not depend on external LLM calls, ensuring stable and deterministic performance. ---

🚀 Setup & Run

Local

pip install -r requirements.txt
python baseline.py      # verify logic
python server.py        # start API server

Visit: http://localhost:7860/docs

Local app entrypoint

python main.py

Docker

docker build -t business-strategy-env .
docker run -p 7860:7860 business-strategy-env

Inference (Rule-Based Agent)

python inference.py

📋 Quick Example

import requests

BASE = "https://ihere04u-business-strategy-env.hf.space"

# Reset
state = requests.post(f"{BASE}/reset", json={"task": "survive", "seed": 42}).json()

# Play 4 quarters
for _ in range(4):
    state = requests.post(f"{BASE}/step", json={
        "task": "survive",
        "action": "increase_marketing",
        "amount": 5000
    }).json()
    print(state["message"], "| Reward:", state["reward"])

# Grade
score = requests.post(f"{BASE}/grader", json={"task": "survive"}).json()
print("Final score:", score["score"])

📁 Project Structure

business-strategy-env/
├── environment.py     # Core simulation logic + stochastic market dynamics
├── graders.py         # Task-specific graders returning scores in [0.0, 1.0]
├── server.py          # FastAPI server — all OpenEnv + additional endpoints
├── baseline.py        # Rule-based baseline agent
├── inference.py       # LLM agent using OpenAI-compatible client
├── openenv.yaml       # OpenEnv spec
├── Dockerfile         # Container — deploys on HF Spaces (port 7860)
├── requirements.txt
└── README.md