Using LLMs to Investigate And Reproduce Complex Bugs

This blog post is based on a presentation I recently gave to our engineering team at my day job. I wanted to share more broadly, so I used Gemini to do the slides → blog post translation, then I heavily edited it to make it my own voice and strip it of anything specific to my company. Much of this is specifically about Claude Code. Using LLMs to Investigate And Reproduce Complex Bugs With Reasonable Token Spend Every engineer knows the feeling of staring down a cryptic stack trace buried deep within a complex application framework. Historically, resolving these issues meant hours of tracing nested logic and reading through the code. LLM agents have changed how I approach that work entirely. I am convinced that one of the best places to start with agentic engineering is with RCAs and bug reproduction, because it's verifiable. If Claude Code can produce the exact same stack trace with a failing unit test, or it can hit our APIs with the magic request that triggers the same failing…