Development of LLM-based Agents within the Amazon Ecosystem

This thesis investigates the development of large-language-model-based agents within the Amazon ecosystem, with particular focus on the architectural constraints imposed by enterprise environments. The work first reviews the foundations of LLM agents, including reasoning loops, tool calling and the Model Context Protocol (MCP), and then presents the design and implementation of a prototype agent built on AWS. The proposed system uses Amazon Bedrock as the model backend and a serverless orchestration layer based on AWS Lambda and API Gateway, while introducing a browser-mediated connection and a local proxy daemon to enable secure access to local or credential-bound MCP tools that cannot be directly exposed to the cloud. To assess the approach, the thesis evaluates a reconstructed version of the architecture on a benchmark of realistic sandbox tasks and compares it against classic function-calling and a Strands-based implementation. The evaluation considers answer quality, tool-use behaviour, latency, token usage, and estimated cost over repeated runs. The results show that the proposed architecture is technically viable and remains valuable in environments where direct cloud access to internal tools is not possible. However, the current prototype trails the strongest baseline in task success and factual accuracy, indicating that the main limitation lies not in the architectural idea itself but in the maturity of the orchestration loop and recovery mechanisms. These findings highlight both the practicality of MCP-based enterprise agent design and the need for stronger orchestration, evaluation, and authentication strategies in future iterations.