Mingbird: small open models fail the harness, not the weights

Hao Wang and Ting Huang argue that 2–9B open models on laptops often produce the right next action but fail inside cloud-scale harnesses: tool prefill crowds context, self-correction diverges, tool demos loop, and tasks are silently abandoned. Mingbird is a local-first Windows/Ollama harness whose ten mechanisms map failure form → fix—three headlines: a byte-level net-zero prefill budget (factory prefill pinned ~797 tokens, CI regression), a finish gate that re-reads the original task before accepting completion, and signature-level loop detection (tool + normalized args; escalate at ≥6 identical / ≥15 empty).

On LRAB—4 harnesses × 4 models (2B–35B) × 18 real tasks, 288 cells, one machine, deterministic artifact scoring—Mingbird scores 0.886 overall against 0.631 (goose), 0.479 (opencode), and 0.405 (agent-mini). It is the only harness without a small-model cliff (2B: 0.821 vs 0.017–0.271). Externally on τ²-bench (278 tasks, one protocol), Mingbird totals 0.856 vs 0.791 (native agent) and 0.737 (opencode). A frontier-model probe on the same 18 tasks spans 0.997–0.478 across harnesses; well-formed scaffolds stay within 0.072 of each other. Ablations are reported as directional (same-night replication noise up to 0.069).

Why it matters for code review

Overnight agent factories that FOMO for “more model” may be measuring harness mismatch: the same small weights complete under Mingbird-shaped guards and collapse under cloud prefills and silent finishes. Review local and edge agent PRs for net-zero tool surfaces, completion that re-checks the task text, and loop signatures before another model swap. Continuous verified momentum compounds when the scaffold stops false dones and empty loops—pair with Li on model–harness fit rankings that reverse and Bounded Loops’ verified completion when the gate must separate scaffold defects from model ceilings.