Benzi coding agent reads less source code, hits 78.2% on SWE-bench for under 10 cents a fix
DonkeyTheKing · reddit · 2026-08-28
- Existing coding agents pull code snippets or build embeddings, blowing up token counts and losing context after compaction.
- Benzi's approach: avoid reading source code in the first place, supplying deterministic intelligence via tool calls — e.g. "what functions feed this one" — with the compiler reporting blast radius before and after edits plus static analysis checks.
- Three truth tiers prevent misleading the model: RESOLVED (verified), CANDIDATE (unresolved by static analysis), OBSERVED (actual runtime behavior).
- It read only 9,125 lines to finish tasks vs Claude Code (20,704), DeepSeek's harness (43,598) and OpenCode (65K+, disqualified after repeated failures).
- Scores 78.2% on SWE-bench Verified at under 10¢ per fix; supports Python/JS/TS/Java/C#/C++/C/Go/Rust/Ruby, with a runtime tracer, mid-task model self-upgrades and more.
Related event: Benzi coding agent reads less source, hits 78.2% on SWE-bench(2 posts)→
More from coding & agent
- Agent writes custom NPU kernels in Chinese to maximize low-end compute — yacineMTB · 2026-08-28
- AI Agents Fail in Companies Because Workflows Are Messy, Not Models — serpratik · 2026-08-28
- Developer Argues Agent Memory Should Rely on Schema Over Semantic Retrieval — arpit_bhayani · 2026-08-28
- Driving Gamma via Claude (MCP): Practical Configs to Avoid AI Clichés — cloudybrain07 · 2026-08-28
- Kimi Code 0.39.0 ships experimental Remote Control for accessing local web sessions remotely — KimiDevs · 2026-08-28
- Analysis confirms ExploitGym scorer rejects poisoned agents — moyix · 2026-08-28