DAIRA: Dynamic Analysis Enhances Issue Resolution in Coding Agents
Debugging is the place where coding agents most visibly break down. The typical agent loop — read code, guess a fix, run tests, repeat — compounds tokens and time on complex bugs because the agent is reasoning about what the code should do rather than what it actually does at runtime. A new research framework called DAIRA (Dynamic Analysis-enhanced Issue Resolution Agent) proposes a fundamental architectural fix: embed lightweight runtime monitors directly into the agent's reasoning cycle, so it can query live variable mutations, call stacks, and execution traces before attempting any repair.
The key insight behind DAIRA is that a significant class of real-world bugs — implicit type degradations, polymorphic control flows, state-dependent failures — are simply invisible to static analysis. They only reveal themselves when the code runs. DAIRA's Test Tracing-Driven methodology ensures the agent's next action is always informed by what actually happened at runtime, not by what the code appears to say on paper. The result is a structured semantic report that the agent can reason over, drastically reducing the trial-and-error cycles that waste both time and context window on hard bugs.
Benchmarks on SWE-Bench Verified show substantial improvements over static-only baselines, particularly on the harder bug categories where traditional agents struggle most. For any team building or integrating production debugging agents today, DAIRA represents a likely inflection point: the shift from static-only to dynamic-enhanced agent reasoning isn't a research curiosity — it's the pattern that will define the next generation of reliable coding agents as codebases grow and bugs grow subtler with them.