How to Make Your Coding Agent Look Like an Idiot (capocasa.dev)

A recent experience highlighted the pitfalls of using AI coding agents for complex software projects. Initially, the process of AI-assisted coding seemed seamless, as the agent effectively generated code and tests. However, as the project evolved, the agent began to struggle with making meaningful changes. The author realized that the issue stemmed from their initial problem definition, which assumed the agents could seamlessly handle sweeping code alterations while maintaining existing tests and styles. This situation underscores a crucial lesson for the AI/ML community: while coding agents like GLM and Codex excel at executing precise tasks within well-defined parameters, they falter in exploratory or refactoring scenarios. The author found that successful architectural changes required explicit instructions to change or discard tests, reflecting a need for a clearer, human-like strategy when working with AI in complex coding environments. This insight emphasizes that while automated tools are powerful, their effectiveness is contingent upon how tasks are framed and communicated by human programmers.