Hybrid text-vision coding agents cut token use by 26% in repo tasks

Current coding agents powered by large language models (LLMs) treat entire code repositories as flat text, missing the visual cues—folder hierarchies, dependency graphs—that human developers use to navigate large codebases. A new paper (arXiv:2606.14061, accepted at ASE 2026) presents the first systematic empirical study of whether multimodal LLMs can benefit from visual representations of repositories. The authors evaluated four recent multimodal models on repository-level issue resolution tasks.

Their results reveal a clear trade-off: a strictly vision-only setup actually degrades accuracy and increases token cost, because the lack of symbolic detail forces agents into repeated visual queries. In contrast, integrating visual graphs of repository structure as a supplementary modality alongside standard text interfaces helps agents understand structure more efficiently. Input token consumption dropped by up to 26% while issue-resolution accuracy was maintained or improved. The visualization proved most valuable during fault localization and when the agent autonomously controls its exploration depth. These findings suggest a practical hybrid text-and-vision design for next-generation coding agents, offering a path to more efficient and capable AI-assisted software engineering.