A new experimental coding agent named Rufous has hit Hacker News, offering a specialized harness designed to tame the erratic behavior of small language models. Written in Rust and licensed under Apache-2.0, Rufous targets models in the 2B to 12B parameter range, addressing the common failure mode where these models report success despite failing to execute file operations. The project builds on research from the Mingbird team, introducing a verification layer that checks model outputs against an immutable record of actual tool calls.

Verifying the Unverifiable

The core innovation of Rufous is its "Record" system, which logs every tool call, including declined writes and exit codes, in a format the model cannot edit. Before a turn ends, Rufous compares the model’s self-reported status against this external record. If a model claims to have written a file that never appeared, or if it forgets a previous refusal, Rufous places a single "Hold" on the reply, listing the missing paths. This mechanism prevents the model from looping indefinitely, as an honest "I couldn’t" always passes through the gate.

Safety Rings and Progressive Disclosure

Rufous assumes the user is running a laptop, not a data center, so it starts with a minimal prefill of roughly 1,250 tokens containing just nine core tools. Additional capabilities like web search and memory are disabled by default, requiring the model to explicitly enable them via enable_tools. Every action passes through a five-ring safety gate that enforces reversible operations, denies irreversible system commands, and restricts sub-agents to their own folders. This architecture ensures that even when a small model makes a mistake, the damage is contained and auditable.

Early Performance Metrics

Initial benchmarks conducted on qwen3.5:4b and gemma4:e4b show significant improvements in task completion rates when using Rufous compared to standard agent loops. For a four-turn task list exercise, the accuracy for qwen3.5:4b jumped from 0.25 to 1.00, while gemma4:e4b improved from 0.58 to 0.83. The tool also demonstrated a 33 out of 36 success rate in forcing models to retry tasks after a denied write, a scenario where small models typically give up or hallucinate success.

The Bottom Line

Rufous proves that you don't need a 70B parameter model to get reliable agentic coding if you build the right scaffolding. It’s a pragmatic, Rust-built solution for the local LLM crowd who are tired of gaslighting from 4B models.

Key Takeaways

  • Rufous is an alpha-stage, Rust-based agent harness for 2B–12B models.
  • It uses an immutable Record to verify tool calls against model claims.
  • Early tests show up to 4x improvement in task completion for small models.
  • Supports Ollama, LM Studio, and any OpenAI-compatible endpoint.