Benzi compiler-backed agent cuts source code reading by 43% while achieving 82.4% on SWE-bench Verified

• Benzi, a code intelligence system, achieves 82.4% on SWE-bench Verified for $12.96 while reading 43% less source code than the baseline mini-swe-agent using DeepSeek v4.1Flash.
• The system offloads deterministic code structure resolution to a local compiler, allowing the AI model to focus cognition on higher-level decisions rather than parsing raw codebases.
• Benzi supports 10 + 3 frontend languages and is model agnostic, utilizing a tier taxonomy (RESOLVED, OBSERVED, etc.) to index repositories once and update incrementally on edits.
• The full agent loop is available via CLI or VS Code extension, while the code map is exposed over MCP for free, though auto-reverts are restricted to the full loop.

Architecture and Performance

• The compiler is lightweight and runs locally, using a runtime tracer to observe edges and resolve code relationships.
• By surrendering structural knowledge to the compiler, the system mitigates the cognitive load imposed by large codebases, improving efficiency without sacrificing accuracy.
• The repository at https://varianttech.net/demo provides a 5-second demo, and https://github.com/oooscoos/Benzi contains SWE-bench results and code generated by Benzi, though the system itself is not open source.