Benzi coding agent reads less source, hits 78.2% on SWE-bench

Open-source project Benzi takes a counterintuitive approach to coding agents: read as little source code as possible. By leveraging compiler-based intelligent localization instead of pulling code snippets or embedding retrieval, it cuts token bloat and reaches 78.2% on SWE-bench.

2026-08-28 ~ 2026-08-28 · 2 related posts