A developer tested nine leading AI models on a real Java integration task, building a server to connect with Eclipse's JDT language server. The goal was to create a working MCP server that could navigate code. The result was a uniform failure across all models, including Gemini 3.0 Flash, ChatGPT, and Claude 4.5 Opus.

The core issue wasn't syntax errors, but a fundamental inability to manage system state and asynchronous protocols. Models either used naive timeouts or generated complex but fragile code that failed silently. This exposed a critical gap: AI can generate text that looks like code, but it cannot reason through the lifecycle of an external system.

The experiment highlights the difference between generating snippets and engineering systems. While AI proves useful for isolated tasks, this test shows its limits in handling real-world complexity. The hype around autonomous coding agents appears premature, as these models lack the causal understanding required for robust software integration.