Researchers have built a system that gives AI coding agents a living map of a codebase instead of a single static snapshot.
The tool, called RepoAtlas, works by pulling a task-relevant slice of a repository's code graph based on both the reported issue and what the agent has explored so far. It then renders that slice as both an image and text, giving the agent two parallel ways to read the same structure. As the agent keeps digging through files, RepoAtlas checks whether its view has gone stale and rebuilds it through what the researchers call a select-project-refresh loop. It requires no extra training - it is a module that plugs into existing coding agents.
On the SWE-bench Verified benchmark, RepoAtlas lifted the resolve rate by 2.4 points over the strongest existing multimodal graph baseline, while using 5.8% fewer input tokens and 7.8% fewer model calls on average. Those savings held across three different model families and sizes, which matters more than the headline accuracy bump: it suggests the approach is not just tuned to one model's quirks.
A 2.4-point improvement will not settle arguments about whether AI can really untangle a legacy codebase, but shaving tokens and API calls while it tries is the kind of unglamorous efficiency gain that actually ships.