Answering Developer Questions with Annotated Agent-Discovered Program Traces [pdf] (andrewhead.info)

Researchers introduced Trailblazer, an IDE tool that answers cross-cutting “reachability” developer questions by combining an AI agent with lightweight program-flow analysis to produce annotated, replayable program traces. The agent simulates an information-forager: it incrementally traces dependencies (using IDE primitives like go-to-definition and rough/static flow analysis), prunes irrelevant paths with AI filtering, and bundles findings into an answer digest — a concise natural-language summary, a descriptive “tour” of key snippets, and a step-by-step annotated walkthrough that can be replayed to show how the answer was discovered. Trailblazer addresses gaps in both traditional flow-tracing tools (which leave the user to navigate many branches) and current LLM-based assistants (which often ignore real code flow). In a within-subject lab study (20 participants), Trailblazer helped developers answer reachability questions more efficiently and improved recall of program flow compared with a modern question-answering baseline. Technically, the system leverages incremental static analysis to create a pruned program tree and uses agent-driven exploration to prioritize promising paths, making explanations verifiable and navigable. This approach makes it easier to understand behaviors spread across files or obscured by callbacks, dynamic dispatch, or layers of indirection—useful for debugging, code review, and refactoring.