This paper presents a modular multi-agent system designed for collaborative problem-solving using LangGraph, a graph-based orchestration framework for language agents. Unlike many existing implementations that depend on proprietary APIs such as OpenAI, this system leverages only free and open-source components, including HuggingFace's flan-t5-base model and SerpAPI for web search.
The architecture comprises three specialized agents—Research Agent, Analyst Agent, and Writer Agent—each responsible for a distinct stage in the reasoning pipeline. The Research Agent gathers real-time information from the web, the Analyst Agent performs logic-driven synthesis using a Python REPL and lightweight LLMs, and the Writer Agent generates coherent summaries of the analyzed data. These agents are orchestrated in a LangGraph workflow that enables modular communication and state transitions.
To demonstrate the system’s practical utility, we simulate a scenario exploring the impact of electric vehicles in India by 2030. The results highlight how free LLMs, when effectively orchestrated in a collaborative agent framework, can approximate high-quality, end-to-end analytical workflows. This project not only showcases the feasibility of cost-free AI agent systems but also opens pathways for scalable, accessible, and transparent multi-agent research tools in education, journalism, and policy analysis.
As the capabilities of large language models (LLMs) advance, the next frontier in artificial intelligence lies in enabling multiple agents to collaborate, each with specialized roles and tools. Multi-agent systems (MAS) simulate how intelligent agents can communicate, share responsibilities, and perform complex reasoning beyond the scope of a single model. These systems mirror real-world workflows where tasks like research, analysis, and report writing are divided across expert roles.
While recent developments in agent orchestration often rely on proprietary services such as OpenAI’s GPT APIs, these solutions introduce limitations in accessibility, transparency, and cost. In contrast, this project explores how a fully open, cost-free, and modular multi-agent system can be built using only publicly available tools such as HuggingFace models, SerpAPI, and LangGraph.
LangGraph—a graph-based orchestration framework—provides a natural and intuitive way to structure agent interactions through state transitions. Each node in the graph represents an agent with a defined function, and edges define the flow of information between them. This framework enables reusability, traceability, and flexibility, making it ideal for academic experimentation and scalable deployments.
This project implements a three-agent system:
The goal is to simulate a collaborative process in which the agents work together to answer high-level questions, such as analyzing the societal and economic impact of electric vehicles in India by 2030. The system’s modularity also allows for easy integration of new tools or domain-specific agents.
This paper details the design, implementation, and demonstration of this system, with a focus on its accessibility, extensibility, and educational value. The work emphasizes the power of open tools and structured agent collaboration in enabling intelligent, autonomous decision-making workflows without reliance on commercial APIs.
The system architecture follows a modular multi-agent design, where each agent performs a distinct function within a defined workflow. The methodology involves the integration of open-source LLMs, task-specific tools, and a stateful orchestration engine (LangGraph) to enable seamless communication and task coordination among agents. The system is implemented in Python and executed via a Jupyter/Colab notebook environment.
The system includes three core agents, each implemented as a callable function that accepts and returns a shared state object:
flan-t5-base): For summarization and insight extraction.flan-t5-base).The interaction between agents is controlled using LangGraph, a graph-based agent orchestration library. The graph structure is defined as follows:
TypedDict called WorkflowState maintains keys like input, research, analysis, and output.LangGraph provides a declarative way to set the entry point, intermediate processing steps, and the final termination of the graph. The shared state is updated at each node, ensuring data integrity and traceability.
| Component | Details |
|---|---|
| LLM | google/flan-t5-base (HuggingFace Transformers) |
| Search | SerpAPI (web search API) |
| Logic/Math | Python REPL (LangChain tool) |
| Orchestration | LangGraph |
| Environment | Google Colab / Jupyter Notebook |
Each step modifies the WorkflowState, and the final result is returned from the LangGraph execution engine.
To evaluate the performance and practical utility of the proposed multi-agent system, we conducted a series of experiments focused on real-world question answering, specifically around socially relevant and research-intensive topics. The goal was to observe the agents' collaborative behavior, effectiveness in task division, and quality of generated output when using only open-source and free tools.
google/flan-t5-base from HuggingFaceEach run began with a natural language query provided by the user, and the agents executed in sequence as per the LangGraph state machine.
A variety of queries were used to test different cognitive and reasoning dimensions of the agents:
Since this is a generative system involving LLMs, qualitative evaluation metrics were prioritized:
| Criterion | Description |
|---|---|
| Relevance | How well the final output aligned with the input query |
| Completeness | Whether the response incorporated multiple perspectives or dimensions |
| Coherence | Logical flow and grammar of the generated report |
| Agent Coordination | Smooth transition of information between agents without data loss |
| Execution Robustness | Error handling, failure recovery, and interpretability of intermediate states |
Each generated report was reviewed manually for these attributes.
flan-t5-base was effective for concise summarization, though longer-context reasoning had some limitations.flan-t5-base is sufficient for summarization but may not scale well for in-depth reasoning compared to larger models.The multi-agent system demonstrated effective collaboration among agents, successfully completing a range of analytical tasks using free tools and open-source models. Below are the results observed during experimental runs with representative queries.
What will be the impact of electric vehicles in India by 2030?
"Electric vehicles are poised to transform India's transportation sector by reducing dependency on fossil fuels and significantly lowering greenhouse gas emissions. Widespread adoption could lead to improved air quality, economic growth in renewable energy sectors, and job creation. However, challenges remain in charging infrastructure development, energy grid readiness, and policy enforcement. By 2030, EVs could represent a major step toward sustainable mobility if investments and regulatory efforts align effectively."
flan-t5-base model.| Evaluation Aspect | Observation |
|---|---|
| Relevance | High – Output directly addressed user prompts |
| Factual Accuracy | Moderate – Depended on up-to-date web results and summarization fidelity |
| Coherence | High – Sentences were logically structured and readable |
| Role Separation | Clear and effective task delegation across agents |
| Execution Robustness | Stable execution in most cases; occasional retries due to SerpAPI timeouts |
| Query Topic | Output Quality | Research Depth | Agent Coordination |
|---|---|---|---|
| Electric Vehicles in India (2030) | ✅ High | ✅ Strong | ✅ Smooth |
| Remote Work in Developing Nations | ✅ High | ✅ Strong | ✅ Smooth |
| Climate Change and Agriculture in South Asia | ✅ High | ✅ Strong | ✅ Smooth |
| AI vs Human Creativity | ✅ Moderate | ✅ Moderate | ✅ Smooth |
flan-t5-base are reliable for short-form summarization but may struggle with multi-hop reasoning tasks.This project demonstrates the feasibility and effectiveness of building a collaborative multi-agent system using only free and open-source tools. By orchestrating specialized agents—each responsible for research, analysis, and report generation—within a LangGraph framework, we created a scalable pipeline capable of addressing complex real-world queries.
Our results show that even without access to premium APIs like OpenAI, it is possible to achieve meaningful and high-quality outputs by combining lightweight LLMs (e.g., flan-t5-base), real-time web search (SerpAPI), and tool-based reasoning (Python REPL). The modular nature of the system ensured clean role separation, easy debugging, and future extensibility, while LangGraph provided robust stateful orchestration.
The project's use case—assessing the impact of electric vehicles in India by 2030—illustrates the system’s potential in research, policy analysis, and education. The experiment validated that free alternatives can yield insightful summaries, actionable conclusions, and well-structured reports when coordinated effectively.
Ultimately, this work underscores a growing opportunity: to democratize AI development through accessible, modular, and transparent multi-agent frameworks—making AI innovation more inclusive and less dependent on proprietary ecosystems.