🤖 Multi-Agent Competitive Intelligence System (CompetiTrack)

An advanced, fully autonomous Multi-Agent Workflow built using LangGraph and Google Gemini, designed to conduct comprehensive competitive intelligence, strategic analysis, and market research.

This project demonstrates the transition from a single-agent architecture to a highly scalable, collaborative multi-agent ecosystem, emphasizing Separation of Concerns, Dynamic Routing, and Production-ready Guardrails.

---

🏗️ System Architecture & Workflow

Unlike traditional linear scripts, this system operates as a virtual "war room" where two specialized AI agents collaborate through a shared persistent state.

1. 🕵️‍♂️ The Researcher Agent:
   • Equipped with the TavilySearch tool.
   • Role: Relentlessly scours the web for raw facts, pricing, and reviews. Strictly forbidden from writing reports.
2. 🧠 The Strategic Analyst Agent:
   • Role: Reads the raw data gathered by the Researcher. Synthesizes market gaps, pricing tiers, and competitive advantages into a professional report.
3. 🔀 The Dynamic Router (Supervisor):
   • Acts as the brain of the workflow. Evaluates the state and creates a Ping-Pong effect: if the Analyst needs more data, the Router sends the workflow back to the Researcher, stopping only when the password (FINAL ANSWER) is issued.

📸 Workflow in Action

Here is a live example of the Ping-Pong communication and the Human-in-the-Loop fallback triggering after reaching the iteration limit:

---

🛠️ Advanced Engineering Features Demonstrated

This project is built with a focus on Agentic System Engineering, applying the following core concepts:

1. State Management & Context Handoff

Utilized LangGraph's StateGraph and TypedDict to create a shared messages memory. Implemented context isolation by dynamically injecting and filtering SystemMessages per node to ensure each agent maintains its distinct persona without cognitive overload.

2. Checkpointing & Memory (MemorySaver)

Integrated langgraph.checkpoint.memory.MemorySaver to provide thread-level persistence. This allows the system to remember its state at any given node, enabling advanced fallback mechanisms.

3. Production Guardrails (Failure Modes Handling)

Configured a strict recursion_limit to prevent infinite tool-calling loops (a common failure mode in LLMs).

4. Human-in-the-Loop (HITL) Fallback Strategy

If the recursion limit is hit (e.g., searching for obscure data), the system cleanly halts via GraphRecursionError, gracefully catches the exception, and actively asks the human user for permission via CLI to generate a partial report using the current checkpoint memory.

5. Safe Output Extraction

Engineered an intelligent output parser to handle Gemini's "Model Turn" payload formats, safely extracting strings from lists to prevent AttributeError crashes during fallback scenarios.

---

📊 Full Observability with LangSmith

Fully integrated with LangSmith for granular tracing of node execution times, token usage, and LLM inputs/outputs. Below are the internal trace trees of the multi-agent execution:

Trace Overview & Output Generation

Internal Fallback Handoff Logic

Final Synthesis & Execution

---

💻 Tech Stack

• Frameworks: LangChain, LangGraph
• LLM: Google Gemini (gemini-1.5-flash) via langchain-google-genai
• Search Engine: Tavily Search API
• Observability: LangSmith

---

🚀 Installation & Setup

1. Clone the repository:

   git clone [https://github.com/YOUR_GITHUB_USERNAME/Multi-Agent-CompetiTrack.git](https://github.com/YOUR_GITHUB_USERNAME/Multi-Agent-CompetiTrack.git)
   cd Multi-Agent-CompetiTrack