ychanakya12 commited on
Commit
5f50076
ยท
1 Parent(s): 0f28872

Detailing Tri-Agent Ecosystem and final documentation fixes

Browse files
Files changed (1) hide show
  1. README.md +40 -53
README.md CHANGED
@@ -30,6 +30,26 @@ The agent leverages **GPT-OSS 20B** to make quarterly decisions across six cruci
30
 
31
  ---
32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33
  ## ๐Ÿง  Problem Statement
34
  The **BusinessSimEnv** is a multi-turn, delayed-reward environment where every action has cascading consequences. A CEO must survive up to **8 quarters** (2 years) while maintaining:
35
  * Positive cash flow in a recession.
@@ -44,76 +64,43 @@ The **AdversarialAgent** in the environment injects realistic "shocks" (budget a
44
  Our agent implements a **Reasoning-Driven Control Loop** with the following advanced heuristics:
45
 
46
  ### ๐Ÿ“Š 1. Proactive ROI Analysis
47
- The agent doesn't just look at base profit. It performs a real-time **Return on Investment (ROI)** analysis that factors in:
48
- * Expected Profit * (1 - Total Risk).
49
- * Quarterly Salary Burn (Salary per developer * count).
50
- * Opportunity Cost of the team's resource pool.
51
 
52
  ### ๐Ÿ“‰ 2. Salary Optimization Hack
53
- To maximize scores on the **Easy (Survival)** and **Medium (Growth)** tasks, the agent implements a "ruthless efficiency" model. It dynamically **fires developers** if the current project portfolio can be managed with a smaller team, instantly saving ~$24,000 per quarter in overhead.
54
 
55
  ### ๐Ÿ›ก๏ธ 3. Bankruptcy Protection
56
- The agent follows a strict **Budget-to-Penalty Ratio**. If a project failure would deplete >50% of the remaining budget, the agent will **"Wait"** and skip the quarter to ensure survival, rather than taking a "suicidal" gamble.
57
-
58
- ### ๐Ÿ’Ž 4. Premium Stack Multiplier
59
- By defaulting to a **Premium Tech Stack**, the agent reduces project risk and increases profit margins by 20%, which is the key to hitting the $150,000 budget target required for a perfect score in the Survival task.
60
 
61
  ---
62
 
63
- ## โš™๏ธ Installation & Usage
64
-
65
- 1. **Clone the Repository**:
66
- ```bash
67
- git clone https://github.com/your-username/business-sim-agent.git
68
- cd business-sim-agent
69
- ```
70
-
71
- 2. **Configure Environment Variables**:
72
- ```bash
73
- $env:HF_TOKEN="your_huggingface_token"
74
- ```
75
-
76
- 3. **Run Inference**:
77
- ```bash
78
- python inference.py
79
- ```
80
-
81
- 4. **Validate Submission**:
82
- ```bash
83
- python validate.py
84
- ```
85
-
86
- ---
87
-
88
- ## ๐ŸŽจ Visualization
89
- The agent provides a clean terminal output with `[START]`, `[STEP]`, and `[END]` markers, as required by the OpenEnv validator, making it fully compatible with automated leaderboard evaluation.
90
-
91
- > *"Building a resilient business is not about never failing; it's about failing small and winning big."* โ€” **AI CEO Agent**
92
 
93
  ---
94
 
95
  ## ๐Ÿ› ๏ธ Tech Stack & Dependencies
96
- The Agent and Environment are built with a modern, high-performance Python stack designed for low-latency inference and robust state management:
97
- * **Core Logic**: Python 3.12+ with strict Pydantic typing for action/observation schemas.
98
- * **Inference**: Leveraging **OpenAI-compatible endpoints** (Hugging Face Router) with **GPT-OSS 20B** for logical reasoning.
99
- * **Environment API**: **FastAPI** for the OpenEnv simulation server, ensuring standard-compliant communication.
100
- * **Communication**: **Httpx** for asynchronous networking between the agent and the simulation.
101
 
102
  ---
103
 
104
- ## ๐Ÿ“Š Evaluation Methodology
105
- Our agent is validated against three distinct "Gates" to ensure it meets Meta's resilience standards:
106
- 1. **Unit Tests**: Verifying deterministic logic for budget calculations and salary burns (`test_grade.py`).
107
- 2. **Adversarial Stress Test**: Running the Hard task (`adversarial_resilience`) which injects 10% chance shocks each quarter.
108
- 3. **Cross-Scenario Generalization**: Ensuring the prompt heuristics work across Easy, Medium, and Hard tasks without manual parameter tuning.
109
 
110
  ---
111
 
112
- ## ๐Ÿ”ฎ Future Roadmap (v2.0)
113
- The next evolution of the MetaStableMinds CEO Agent will focus on:
114
- * **Multi-Agent Negotiation**: Allowing the CEO to negotiate project costs with the Adversarial Agent.
115
- * **Dynamic Tech Stacks**: Implementing custom tech stacks where the agent can mix-and-match technologies.
116
- * **Reinforcement Learning (PPO)**: Transitioning from heuristic prompts to a fine-tuned model specifically trained on BusinessSim trajectories.
117
 
118
  ---
119
 
 
30
 
31
  ---
32
 
33
+ ## ๐ŸŒ The Simulation Ecosystem: A Tri-Agent Interaction
34
+ Our simulation is not a static set of rules, but a dynamic, multi-agent world where the **AI CEO** must compete and survive against two other environmental agents:
35
+
36
+ ### ๐Ÿ‘ค 1. The AI CEO (The Decision Engine)
37
+ * **Role**: The primary protagonist of the simulation.
38
+ * **Actions**: Project acceptance, hiring/firing (Salary Optimization), tech stack selection, and skill training.
39
+ * **Objective**: Maximize long-term corporate health (Budget, Reputation, and Low Burnout) through high-reasoning ROI analysis.
40
+
41
+ ### ๐Ÿ“‰ 2. The Market Agent (The Economic Engine)
42
+ * **Role**: Simulates the macro-economic environment and consumer demand.
43
+ * **Behavior**: Driven by a Markov Chain, this agent triggers **Boom, Stable, and Recession** cycles.
44
+ * **Impact**: It alters the Domain Demand multipliers (e.g., higher demand for AI during Booms) and profit variances, forcing the CEO to adapt their strategy quarterly.
45
+
46
+ ### ๐Ÿ›ก๏ธ 3. The Adversarial Agent (The Friction Engine)
47
+ * **Role**: Simulates the "Chaos" of real-world business operations.
48
+ * **Behavior**: Injects randomized "Shocks" based on the simulation's difficulty level.
49
+ * **Impact**: Triggers events like **Budget Audits** (draining cash), **Key Employee Departures** (lowering skill), and **Client Disputes** (tanking reputation).
50
+
51
+ ---
52
+
53
  ## ๐Ÿง  Problem Statement
54
  The **BusinessSimEnv** is a multi-turn, delayed-reward environment where every action has cascading consequences. A CEO must survive up to **8 quarters** (2 years) while maintaining:
55
  * Positive cash flow in a recession.
 
64
  Our agent implements a **Reasoning-Driven Control Loop** with the following advanced heuristics:
65
 
66
  ### ๐Ÿ“Š 1. Proactive ROI Analysis
67
+ The agent performs real-time **ROI** analysis factoring in Expected Profit, Total Risk, and Quarterly Salary Burn.
 
 
 
68
 
69
  ### ๐Ÿ“‰ 2. Salary Optimization Hack
70
+ The agent dynamically **fires developers** to instantly save ~$24,000 per quarter in overhead.
71
 
72
  ### ๐Ÿ›ก๏ธ 3. Bankruptcy Protection
73
+ Follows a strict **Budget-to-Penalty Ratio**. Skip the quarter rather than taking a "suicidal" gamble.
 
 
 
74
 
75
  ---
76
 
77
+ ## ๐Ÿ“Š Evaluation Methodology
78
+ Our agent is validated against three distinct "Gates":
79
+ 1. **Unit Tests**: Verifying deterministic logic for budget calculations (`test_grade.py`).
80
+ 2. **Adversarial Stress Test**: Running the Hard task (`adversarial_resilience`).
81
+ 3. **Generalization**: Heuristics that work across all task difficulties.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
82
 
83
  ---
84
 
85
  ## ๐Ÿ› ๏ธ Tech Stack & Dependencies
86
+ * **Core Logic**: Python 3.12+ with Pydantic schemas.
87
+ * **Inference**: **GPT-OSS 20B** on Hugging Face Router.
88
+ * **Environment API**: **FastAPI** simulation server.
89
+ * **Communication**: Asynchronous **Httpx**.
 
90
 
91
  ---
92
 
93
+ ## ๐Ÿ”ฎ Future Roadmap (v2.0)
94
+ * **Multi-Agent Negotiation**: Direct negotiation with the Adversarial Agent.
95
+ * **Dynamic Tech Stacks**: Modular technology selection.
96
+ * **Reinforcement Learning (PPO)**: Transitioning from heuristic prompts to fine-tuned models.
 
97
 
98
  ---
99
 
100
+ ## โš™๏ธ Installation & Usage
101
+ 1. **Configure Token**: `$env:HF_TOKEN="your_token"`
102
+ 2. **Run Inference**: `python inference.py`
103
+ 3. **Validate**: `python validate.py`
 
104
 
105
  ---
106