Agentic Software Engineering in Practice

The Project: AI-Native Development of a Port Railway System 

Our practical example is a complex, multi-tenant dispatching and billing system for port railways. The goal was to develop all operations—ranging from graphical scheduling and automated billing to comprehensive system integrations—entirely with AI support. The challenge: no manual coding, high complexity, real-time requirements, and scalable teamwork. 

From Planning to Implementation: When AI Takes Over Coding 

The development team started with a clear rule: no one is allowed to type code—everything is done via prompts and AI agents. After initial experiments with various tools, the “Cursor AI” editor proved particularly convincing. Acting as an agent, it gained access to the file system, created files autonomously, wrote code, and gradually built up the project piece by piece. It quickly became apparent: for individuals, this works surprisingly well, but in a team, new challenges arise. The agent only knows its own context and is unaware of changes made by other team members. This leads to conflicts and increases the need for coordination. 

Context Is Everything: The “Memory Bank” as Knowledge Repository 

To avoid loss of context between work sessions, a so-called “Memory Bank” was introduced—a structured knowledge base that documents tasks, context, and design decisions. This allowed the agent to be quickly brought up to speed each morning. The focus shifted: 80% of the work was planning, only 20% was actual implementation—and most of that was handled by the agent. The key insight: the more precise and complete the planning, the better the results. 

New Roles, New Challenges 

The traditional division of labor in development teams—frontend, backend, QA—is blurring. With agents, everyone becomes a “full stack” developer, taking responsibility for entire features. Testing remains a challenge: while AI can quickly generate tests, humans still need to check whether these tests actually verify the desired behavior. The best results are achieved when the team invests two to three weeks (or more) in planning before letting the AI get started. 

What (Still) Doesn’t Work – and Where We’re Headed 

Anyone who thinks everything can be automated with a single prompt will be disappointed. Agentic coding is not a silver bullet. The tools are powerful, but only as good as the prompts and preparation. Discipline in planning and documentation is essential, and AI is not a “magic button” that does everything automatically. The team learned that the best results come from precise prompts, structured knowledge management, and regular reflection. The psychological shift shouldn’t be underestimated either, as the work changes fundamentally: less typing yourself, more planning, more explaining—and that can be more exhausting than expected. Nevertheless, the potential is enormous: complex systems can be realized in record time when humans and AI collaborate effectively. The future? It belongs to teams that combine planning, engineering know-how, and AI intelligently. 

Our Conclusion from the Agentic AI Field Test 

AI is fundamentally changing business engineering and software development. Precision and planning are becoming more important, and the interplay between human and machine is the key to success. While humans remain essential as critical reviewers and creative planners, AI is taking over more and more routine tasks—opening up new opportunities for speed and innovation.