Mingbird is a local-first agent harness for Windows and Ollama designed to help small open-weight models (2-9B) complete real tasks by addressing common failure modes like context overflow and looping. It employs ten mechanisms, including a byte-level net-zero prefill budget and signature-level loop detection, to compensate for these issues.

  • On the LRAB benchmark, Mingbird achieved an overall score of 0.886, outperforming goose (0.631), opencode (0.479), and agent-mini (0.405).
  • In τ²-bench evaluations, it reached a total of 0.856 compared to 0.791 and 0.737 for other harnesses.
  • The system includes a finish gate that re-reads the task before accepting completion to prevent divergence.

The authors argue that a substantial share of failures in small models is attributable to the harness rather than the model itself, demonstrating that Mingbird's approach significantly boosts reliability.