Agentic AI
New LLM Framework Tackles Repository-Wide Code Fixing, Boosting Accuracy by 4%
SWE-Adept splits code analysis into two agent roles to navigate large codebases more efficiently and resolve bugs with version-control precision.
Large language models have become adept at isolated coding challenges, but struggle when tasked with real-world software engineering across entire repositories. A new framework called SWE-Adept addresses this gap by breaking the problem into two specialized agent roles that work together to find and fix issues at scale.
The Localization-Resolution Pipeline#
The framework deploys a pair of agents with distinct responsibilities. The first—a localization agent—pinpoints the code sections most relevant to a reported issue, sifting through vast dependencies to avoid cluttering the model's context window. The second—a resolution agent—then implements the actual fixes. This division of labor mirrors how experienced developers approach unfamiliar codebases: identify the problem zone first, then systematically repair it.
Smart Dependency Traversal#
To locate problematic code efficiently, SWE-Adept uses agent-directed depth-first search, a technique that selectively walks through code relationships rather than loading everything at once. This keeps the LLM's limited context focused on what matters, improving the odds of pinpointing the right location and reducing irrelevant noise.
Structured Fixing with Version Control#
For the resolution phase, the framework couples the agent with specialized tools that track progress and manage Git operations. A shared working memory stores snapshots of the code state indexed by execution step—think of it as a map of checkpoint locations. This design enables the agent to branch experiments and revert failed attempts without manual overhead, supporting a methodical trial-and-error workflow.
Results on Real Benchmarks#
Testing on SWE-Bench Lite and SWE-Bench Pro, which measure end-to-end issue resolution across real repositories, SWE-Adept improved performance by up to 4.3% over prior approaches. The gains come from combining better localization accuracy with more reliable, version-control-enabled fixing strategies.
