Show HN: Try Benzi – A coding harness/agent beating Claude Code itself on Sonnet

What happened

A developer going by the handle on Hacker News has released Benzi, a coding harness and agent system that claims to outperform Claude Code on its own benchmark suite. The creator posted under "Show HN," sharing a public benchmark page at benzi.fly.dev/benchmark, and the headline finding is stark: on 2 out of 20 tests, Claude Code — running primarily on Sonnet — failed where Benzi succeeded. The project is framed as something the creator believes "should've been made a long time ago," suggesting frustration with how existing agents handle codebase comprehension at scale rather than novelty for its own sake.

The core architectural idea behind Benzi is a departure from how most coding agents navigate code. Rather than streaming file contents into context windows or relying on fuzzy semantic search over embeddings, Benzi compiles the entire codebase into O(1) hashmaps. The agent queries these hashmaps directly to discover code structure, answer questions about the codebase, and generate new code. This means lookup time for any structural fact about the codebase is theoretically constant regardless of project size, which is a meaningful departure from approaches that degrade as repositories grow. On top of that, every write the agent proposes goes through complete static analysis before it lands, giving the system a verification layer that most agent frameworks lack entirely.

The technical detail

For engineers building on top of LLMs, the hashmap approach addresses a real and persistent problem. Large language models have bounded context windows, and even models with extended context struggle with precise retrieval across hundreds of files. Tools that stuff raw code into prompts either truncate aggressively or blow through token budgets fast. Benzi's approach sidesteps this by pre-indexing structure so the agent queries a fast lookup layer rather than reasoning over raw text. The static analysis pass on writes is arguably just as important — it catches ill-formed edits before they propagate, which is the single most common failure mode when agents autonomously modify real codebases.

Why it matters for builders

The competitive framing here is deliberate and pointed. Claude Code is Anthropic's own agentic coding product, built specifically around the Sonnet model family, and positioned as a serious tool for professional software development. Beating it on a benchmark, even partially, on its own model is a direct challenge to the premise that tightly integrated first-party tooling has inherent advantages. The benchmark claims are specific: 2 of 20 tests where Claude Code failed but Benzi passed. The creator notes Claude Code was "mostly Sonnet," which is careful phrasing — it leaves open whether model variation was controlled across all 20 tests, a methodological detail the benchmark page itself would need to clarify.

The practical takeaway for builders right now is to look at the benchmark methodology before drawing strong conclusions. The test suite is public and accessible, which is the right move — it invites independent verification and exposes the claims to scrutiny. What matters most is what the 20 tasks actually represent: are they representative of real-world engineering tasks, or are they optimized for the strengths of Benzi's architecture? Constant-time structural lookup helps enormously with tasks requiring precise symbol resolution and cross-file dependency tracing, but it may matter less for tasks that are primarily generative or require nuanced semantic reasoning about behavior rather than structure.

Market and competitive context

The open questions are significant. Static analysis coverage varies enormously by language, and the creator has not specified which languages Benzi supports or which analyzers back the verification layer. If the system works well for typed, statically analyzed languages like TypeScript, Go, or Rust but struggles with Python or dynamic languages, that narrows the addressable use case considerably. There is also a prior art question worth noting: tools like Sourcegraph's Cody, Aider, and various LSP-backed agent harnesses have all experimented with structured code indexing rather than naive context stuffing. What distinguishes Benzi's hashmap approach from those existing indexing strategies is not yet fully documented in the public post, and the technical depth of the explanation will likely determine whether this attracts serious engineering attention or reads as a well-marketed prototype.

What to watch next

The Hacker News framing — "Show HN" rather than a formal launch — keeps expectations appropriately calibrated. This is early-stage, creator-shared work, not a product announcement from a funded company. The benchmark link is the real asset here. If the methodology holds up to community scrutiny and the task set is genuinely representative, Benzi represents a useful existence proof that purpose-built harness architecture can compensate for gaps in raw model capability. If the tasks are narrow or the comparison is not apples-to-apples across model invocations, the claims deflate quickly. Either outcome would be informative, which is precisely why publishing the benchmark publicly was the right call.