Design and Implementation of a Custom Multi-Agent Platform for LLM-based AWS Cloud Orchestration

The thesis addresses the limitations of Large Language Models (LLMs) in handling long-term, complex tasks such as maintaining contextual continuity and interacting with external systems and APIs. While LLMs have shown remarkable capabilities in natural language understanding and generation, they often struggle with multi-step reasoning, dependency management, and robust integration with cloud platforms. To overcome these challenges, the project proposes the design and implementation of a custom Multi-Agent System (MAS), where LLM-based agents work collaboratively to autonomously orchestrate AWS cloud services in response to natural language instructions. The platform, developed in collaboration with Data Reply, aims to support software developers throughout the Software Development Life Cycle (SDLC) by automating tasks such as cloud infrastructure setup, service configuration, deployment, and monitoring. The system architecture is built around three core agents: a planner that decomposes user instructions into executable tasks such as cloud infrastructure setup, service configuration, deployment, and monitoring. The system architecture is built around three core agents: a planner that decompose user instructions into executable tasks, an extractor that retrieves and validates required parameters, and an executor that performs the actions using the Boto3 SDK. The system is designed to follow a structured reasoning principle through Pydantic schemas and a lightweight but effective contextual memory mechanism ensures continuity during user sessions, enabling the system to disambiguate references and maintain coherence across multi-step workflows. Tools like LangChain are used to interface with LLMs, supporting structured outputs, memory management, and external tool invocation. Together with Pydantic and Boto3, these framrworks form a technological ecosystem that balances innovation with robustness, making the prototype both extensible and production-oriented. Logging and cost monitoring mechanisms were integrated to track AWS resource usage, LLM token consumption, and execution traces, ensuring transparency, accountability, and enterprise readiness. Security was addressed through credential management policies, enforcing that AWS keys remain under the exclusive control of the user. The architecture has been rigorously tested using a suite of 24 realistic scenarios covering Lambda deployment, API Gateway integration, EventBridge scheduling, DynamoDB operations, and EC2 management. Results demonstrated a success rate of 87.5%, with only a few failures linked to ambiguous prompts or complex integrations. Beyond quantitative outcomes, the tests confirmed the system ability to handle high-level natural language instructions , reuse conversational context, and gracefully recover from incomplete specifications. These insights highlight not only the technical feasibility of the approach but also its cost-effectiveness and reliability for enterprise contexts.