Benzi: coding agent that reads less source, hits 78.2% SWE-bench under 10¢ a fix
DonkeyTheKing · reddit · 2026-08-28
Benzi is a coding agent built to avoid reading source code in the first place. Instead of pulling snippets or embeddings — which balloon tokens and drift context — it feeds the model deterministic intelligence via tool calls: before an edit it can query "what functions feed this one?", and the compiler reports blast radius plus static analysis checks.
Correctness is handled with 3 truth tiers: RESOLVED (definite evidence), CANDIDATE (unresolved by static analysis), and OBSERVED (actual execution). Benzi+Sonnet read only 9,125 lines of source vs 20,704 for Claude Code, 43,598 for DeepSeek's harness, and 65K+ for OpenCode. It scores 78.2% on SWE-bench Verified at under 10¢ a fix using V4flash, and supports Python, JS/TS, Java, C#, C++, C, Go, Rust and Ruby, plus extras like runtime tracing and mid-task self-aware model upgrades.
More from coding & agent
- Google Cloud Run instances run long-lived agents for $5.70/month — IanAndrewsDC · 2026-08-28
- Opinion: Models and Harnesses are intertwined, engineering skills are key — omarsar0 · 2026-08-28
- Stanford launches Terminal-Bench-Science: Claude Opus 5 solves only ~30% — ajratner · 2026-08-28
- Don't black-box the harness layer: open-source agent frameworks enable customization — omarsar0 · 2026-08-28
- Dev Opinion: 'Agents' are just reprompts, not new AI lives — gerardsans · 2026-08-28
- Embrace Open-Source Harnesses for Customizable AI Applications — omarsar0 · 2026-08-28