I recently tried to use Serena (an AI coding agent + language server toolkit) and hit a crash during project indexing. Instead of filing a vague issue, I set up Claude Code, indexed Serena’s own codebase using Serena itself, and then asked the AI agent why Serena was failing.
This led to an interesting discovery:
• My initial guess (“large files causing crashes”) was wrong.
• The actual problem was that Serena’s language servers struggled with complex regex/iterator-heavy files.
• I was able to file a much better issue report, with logs and a hypothesis, and propose an improvement (exclude certain files from indexing).
This seems like a useful workflow for open-source contributors: use AI-assisted analysis responsibly before submitting issues, to save maintainers time and avoid noisy back-and-forth.
Here’s the write-up:
Why Half-Baked Bug Reports Waste Everyone’s Time (and How AI Can Level Up Your Open-Source Contributions) (replace with final blog link)
Curious if others are using Claude/Serena for similar “meta-debugging” workflows. Would love to hear examples.
After months of conversations with experienced Magento 2 developers about AI tooling, the feedback is brutal: "AI delivers horrible code," "It's fundamentally stupid," "We tried it and went back to writing everything ourselves."
They're not wrong about their experience—they're wrong about the cause.
Recent METR research on AI-experienced developers found productivity reductions, not gains, when using AI tools for complex tasks. The problem isn't AI capability. It's that most developers treat AI agents like junior developers: throw requirements over the wall and expect clean code back.
This approach fails catastrophically because it ignores fundamental limitations like the "Lost in the Middle" problem, where LLMs degrade performance on long contexts, and error propagation in multi-step reasoning.
The implementations that work treat AI agents as specialized components in larger systems, with proper validation, error handling, and human oversight. Success comes from better architecture, not better prompts.
*Full technical breakdown of production-tested patterns inside.