Interpreting Coding Agents Through Their Behavioral Trajectories
Researchers analyzed a large corpus of over 143,000 execution trajectories across various agent configurations and models to understand how coding agents solve tasks. They developed a rule-based parser mapping raw traces to high-level actions and found that agents with identical solve rates exhibit distinct exploration, modification, and verification behaviors. The study demonstrates that aggregate behavioral profiles can successfully retrieve corresponding agent stacks and that scaffold choice and reasoning effort systematically shift agent performance. Open article →