Approaches to building agentic systems based on large language models have been developing rapidly, enabling autonomous agent behaviour through the use of tools, memory, and orchestration. This thesis presents three approaches to developing such systems across three abstraction levels: n8n as a visual low-code/no-code environment, LangChain/LangGraph as a fully customisable open-source framework, and Google ADK as a cloud-oriented solution for production environments. Special attention is given to information security. We present the most significant security risks according to the OWASP Top 10 for LLM Applications 2025 and examine which protective mechanisms are available to developers. The thesis presents an analysis of the three frameworks and three working prototypes: a fake online shop detector (n8n), a CV and cover letter generation tool (LangGraph), and a personal Airbnb assistant (Google ADK). The frameworks are analysed in terms of functionality, implementation complexity, security capabilities, and suitability for production environments. The analysis shows that n8n is best suited for rapid prototyping and visual construction of agentic workflows without deep programming knowledge, that LangGraph provides the greatest control over state and agent orchestration in complex scenarios, and that Google ADK is the optimal choice for production deployment within the Google Cloud environment. A key finding is also that none of the frameworks is secure by default and that developers must implement security mechanisms thoughtfully and systematically, without relying on built-in safeguards.
|