AI Agents - Flomny
Degree
Bachelor of Science (Computer Science)
Department
Department of Computer Science
School
School of Mathematics and Computer Science (SMCS)
Advisor
Dr. Tahir Syed, Assistant Professor, Institute of Business Administration Karachi
Keywords
Multi-Agent Systems, Automation, LLM, DevOps, Software Engineering, Model Context Protocol
Abstract
Traditional workflow automation platforms such as Zapier, IFTTT, and n8n have empowered users to connect disparate services through triggers and actions. However, they typically require either an understanding of API mechanics or the use of rigid graphical interfaces, limiting their accessibility for non-technical users and restricting flexibility for more advanced automation. In parallel, the emergence of Large Language Models (LLMs) has revolutionized how systems can interpret and respond to natural language, offering a unique opportunity to simplify software development and workflow automation through conversational interfaces. Our methodology emphasizes modularity, robustness, and traceability. The system begins with prompt validation and task decomposition using LLMs, followed by graph-based orches- tration of agents. Integration Agents leverage Pinecone for vector-based retrieval and produce code grounded in documentation rather than relying solely on the LLM’s internal knowledge. Additionally, Flomny integrates the Model Context Protocol (MCP) to enable advanced agent- based execution capabilities, allowing for dynamic discovery and utilization of external tools while maintaining security boundaries and user control. Static code validation precedes optional execution in isolated subprocesses, enhancing both safety and reliability. Through a detailed case study involving Custom Servers and Discord integrations, we demonstrate Flomny’s ability to automatically generate and execute a four-step workflow derived entirely from a natural language prompt. The results show successful task decomposition, accurate API calls, and seamless execution with real-time updates. By combining the interpretability of LLMs, the structure of graph-based multi-agent systems, the precision of retrieval-augmented code generation, and the flexibility of MCP integration, Flomny paves the way for more accessible, scalable, and human-centric workflow automation.
Tools and Technologies Used
Tools and Technologies Used
Programming Languages: Golang, Python, TypeScript / JavaScript
Frontend: Next.js
Backend: Gin-Gonic, gRPC, Redis-Events, FastAPI (Websocket), LangChain / LangGraph
Datastores: MongoDB, PineCone, S3
Infrastructure: Docker containers, AWS EKS (Kubernetes based deployments), AWS ECR (Elastic Container Service)
Methodology
The methodology of the Flomny system follows a sophisticated multi-layered approach that transforms natural language prompts into executable automation workflows. At its core, the system employs a two-tiered architecture separating static Integrations (reusable knowledge bases for APIs) from dynamic Workflows (user-generated automation sequences). When a user provides a natural language prompt, it first undergoes validation to ensure it contains appropriate integrations and adheres to content guidelines. The validated prompt is then decomposed by a Task Breakdown Agent into atomic, logical subtasks, which are subsequently refined into structured JSON representations by a Refinement Agent. A Manager Agent orchestrates the workflow by assigning each subtask to specialized Integration Coder Agents that generate executable code using two distinct strategies: for custom integrations, they employ Retrieval-Augmented Generation (RAG) against API documentation stored in a Pinecone vector database using a novel summarization-based chunking approach that improved retrieval precision by 40%; for internet-based integrations, they perform real-time web searches to fetch current documentation. The generated code undergoes rigorous validation through an LLM-based Code Validation Node before being merged into a cohesive workflow script with templated parameters. The system also incorporates the Model Context Protocol (MCP) for advanced agent-based execution, enabling dynamic discovery of external tools and real-time adaptation through hierarchical agent structures. Throughout the entire process, a WebSocket-based communication layer provides real-time progress monitoring, while comprehensive error handling ensures graceful recovery from failures, making the system both robust and user-friendly for non-technical users seeking to automate complex multi-service workflows.
Document Type
Restricted Access
Submission Type
BSCS Final Year Project
Recommended Citation
Ammar, M., Shamsi, M., Masood, R., Thebo, M., & Shehzad, M. (2025). AI Agents - Flomny. Retrieved from https://ir.iba.edu.pk/fyp-bscs/18
