Coding Agents Need Minimal Context for Code Editing.

This research investigates the precise amount and type of context that coding agents truly require to perform code editing tasks effectively. The study differentiates between the act of finding the work site and acting on it, using an oracle to fix localization and focusing solely on how code representation impacts an agent's ability to resolve issues on the SWE-bench Verified benchmark. The findings reveal that coding agents need remarkably little context. The most critical signal is embedded directly within the code being edited. Natural language summaries of the code proved largely ineffective, performing as poorly as summaries from smaller models, and providing almost no behavioral insights compared to the source code itself. Furthermore, surrounding context from other files, even when represented as UML skeletons and signatures, offered no significant improvement over simply deleting that remainder. This suggests that extensive multi-file context is often superfluous for the act of editing. Compressed context, however, matched the performance of whole files using only a third of the tokens, indicating efficient context representation is key. The study also noted a significant noise floor in API inference at temperature-0, causing outcomes to flip in ~9% of runs, which has implications for benchmark reliability.