Releases·Agent Release
Compiler backed agent Benzi beats Claude Code, OpenCode & other major agents and harnesses (benchmarks linked)
A developer shared Benzi, a coding agent built on a compiler backend that gives models deterministic static analysis via tool calls instead of reading large amounts of source code. Its benchmark results reportedly show fewer lines read, lower cost, and faster task completion than Claude Code, OpenCode, and DeepSeek's harness. The project is open on GitHub with a demo and benchmark site.
32 engagement2 sourcesThu, Oct 1, 2026, 09:36 PM
AI Overview
- Benzi avoids reading source code, supplying deterministic compiler analysis via tool calls; 78.2% SWE-bench Verified
- Benzi Sonnet reads 9,125 lines vs Claude Code Sonnet's 20,704 and DeepSeek harness's 43,598; OpenCode (65K+) disqualified for repeated failure
- Supports Python, JS, TS, Java, C#, C++, C, Go, Rust, Ruby, plus HTML/CSS resolution; includes runtime tracer and verified writes
Related
Releases·Agent Release
Benzi harness hits 78.2% SWE-bench Verified while reading less code than Claude Code or DeepSeek
A developer released Benzi, an AI coding harness that supplies models deterministic code intelligence via compiler, static analysis, and runtime tracer tool calls instead of feeding them raw source code. Its benchmarks claim fewer lines read and 78.2% SWE-bench Verified versus Claude Code, DeepSeek's harness, and OpenCode.
A developer released Benzi, an AI coding harness that supplies models deterministic code intelligence via compiler, static analysis, and runtime tracer tool calls instead of feeding them raw source code. Its benchmarks claim fewer lines read and 78.2% SWE-bench Verified versus Claude Code, DeepSeek's harness, and OpenCode.
0 engagement1 sourceredditWed, Sep 16, 2026, 03:11 PM
Products·building
Benzi - compiler-backed code intelligence infrastructure for AI coding agents (CLI/MCP/Harness)
Benzi provides compiler-based code intelligence to AI coding agents instead of text snippets or embeddings, aiming to reduce token counts, wall-clock time, and context drift. It is offered as a harness, CLI, and MCP server. It targets developers building or using AI coding agents on large codebases.
Benzi provides compiler-based code intelligence to AI coding agents instead of text snippets or embeddings, aiming to reduce token counts, wall-clock time, and context drift. It is offered as a harness, CLI, and MCP server. It targets developers building or using AI coding agents on large codebases.
1 engagement1 sourceredditMon, Sep 28, 2026, 11:37 PM
Developments·Paper
What Does a Harness Buy? Tokens, Mostly
An arXiv study runs five models through three production coding-agent harnesses (Claude Code, mini-SWE-agent, OpenCode) on SWE-bench Verified to isolate how much the harness itself moves results. It finds harness-swapping effects are mostly indistinguishable from rerun noise, while cost per task differs up to 3x across harnesses. Practitioners should treat small leaderboard gaps skeptically and weigh harness token overhead.
An arXiv study runs five models through three production coding-agent harnesses (Claude Code, mini-SWE-agent, OpenCode) on SWE-bench Verified to isolate how much the harness itself moves results. It finds harness-swapping effects are mostly indistinguishable from rerun noise, while cost per task differs up to 3x across harnesses. Practitioners should treat small leaderboard gaps skeptically and weigh harness token overhead.
0 engagement1 sourcearxivSat, Oct 3, 2026, 10:47 AM
Developments·Benchmark
I made an open-source benchmark which measures the complexity of code added (overengineering/tech debt). Looking for feedback and contributors.
A developer released Parsimony, an open-source benchmark that measures how much code AI coding agents add, addressing overengineering and tech debt. It reuses published patches from existing benchmarks and counts net "coding units" rather than lines of code. The creator is seeking feedback and contributors.
A developer released Parsimony, an open-source benchmark that measures how much code AI coding agents add, addressing overengineering and tech debt. It reuses published patches from existing benchmarks and counts net "coding units" rather than lines of code. The creator is seeking feedback and contributors.
1 engagement1 sourceredditSat, Oct 10, 2026, 05:14 PM
Products·launched
AI coding assistant grounded in a live code map
Benzi is an AI coding assistant that uses a compiler, runtime tracer, and harness to build a live code map, aiming to understand code more deeply than traditional RAG or embedding approaches. It helps developers write and understand code more cleanly by grounding the AI in actual program behavior.
Benzi is an AI coding assistant that uses a compiler, runtime tracer, and harness to build a live code map, aiming to understand code more deeply than traditional RAG or embedding approaches. It helps developers write and understand code more cleanly by grounding the AI in actual program behavior.
15 engagement2 sourceshackernews, redditTue, Aug 4, 2026, 03:30 AM