How One AI Coding Agent Kept Failing — and the Code Written to Fix It

Software firm Univoco built a retrieval-augmented AI coding agent for a proprietary document layout engine that has no public internet presence, forcing the agent to rely entirely on a private documentation index. Because the engine is absent from the model's training data, the setup offered a rare opportunity to observe pure problem-solving behaviour — and its failures. The agent repeatedly made unexpected errors, such as failing to locate an existing file by searching its contents rather than its filename, or deleting one line of code while leaving related configuration lines intact. Engineers at Univoco documented each failure in the order it occurred and published the specific code fixes applied to resolve them. The post details nine distinct failure modes, covering issues ranging from literal search limitations and infinite search loops to the agent inventing APIs that did not exist.