Two Days of Hackathon - Without Writing a Single Line of Code Ourselves

After the development phase came the debrief: how did coding without writing code actually work? For some, every build felt like a gamble. The same prompt and minimal changes to the documentation often produced a completely different result when generated again. “Suddenly, chess pieces appeared on the results page of our robot comparison, and we had no idea where they came from,” one developer reported.

Philipp Kemper (CHECK24 CTO) had already anticipated this experience in his opening keynote. He described context engineering as the new core competency for developers: deliberately controlling the information available to an AI within its context. Part of this context is static and available with every request, such as global instructions or central rules. Other context is loaded dynamically. Relevant skills, documents, or tool results only enter the context when they are needed for the task at hand. Context engineering therefore determines not only what the model knows, but also when it knows it.

This is exactly where the harness comes into play. An agent consists not only of the model, but of the interaction between the model and the harness. The harness is the technical environment in which the model operates.

This was precisely what made the CHECKathon so interesting: we did not manually adjust either the model or the code. We only changed the documentation, thereby influencing part of the context. The harness, however, remained unchanged and, alongside impressive results, also produced the observed “gambling builds.” This gave our developers first-hand experience of just how important a good harness is. Philipp's recommendation for day-to-day development followed directly from this: “Invest in the harness, not just the model.”