name: business-sim-env version: 0.1.0 description: > Multi-agent business simulation where an AI CEO manages a software company across quarters under adversarial conditions. Tests strategic decision-making with delayed consequences across 6 real-world business factors: profit potential, risk, team capability, market demand, reputation, and resources. tasks: - id: single_quarter_survival difficulty: easy max_quarters: 1 description: End the quarter with positive cash flow. Balance risk vs profit. - id: four_quarter_growth difficulty: medium max_quarters: 4 description: Grow revenue 30% over 4 quarters. Adapt to market phase shifts. - id: adversarial_resilience difficulty: hard max_quarters: 8 description: Survive 8 quarters with reputation >= 0.6 despite adversarial shocks. action_space: CEOAction observation_space: CompanyObservation api: reset: "POST /reset?task_id={task_id}" step: "POST /step?session_id={session_id}" state: "GET /state?session_id={session_id}" tasks: "GET /tasks" grade: "GET /grade?session_id={session_id}" reward: type: continuous range: [-2.0, 2.0] partial_progress: true description: > Shaped reward combining net profit, reputation delta, burnout penalty, and bankruptcy punishment. Partial credit awarded at each step. environment: agents: - name: MarketAgent role: > Simulates dynamic market conditions (boom/stable/recession phases) using a Markov chain. Affects domain demand and project availability. - name: AdversarialAgent role: > Injects realistic business shocks (client disputes, key dev quits, budget audits) to prevent the CEO agent learning brittle strategies. - name: DecisionAgent role: > The LLM-based CEO agent. Makes quarterly decisions via step() API. Trained/evaluated by interacting with the environment.