Trial by Trace - DeepSeek Harness - Agent Pass: leaderboard

Trial by Trace measures model task performance under a named native agent harness on 95 mechanism-focused tasks per model. Agent Pass requires all agent task criteria to pass; Agent Score is criterion-level performance. Harness-only scores, costs and mixed Agent/Harness composites are not admitted. arXiv v1; 95 tasks per model/harness; one selected execution including recovery replacements; DeepSeek-V4-Flash semantic judge with variable scoring budgets; native harness versions, model API snapshots and reasoning effort unreported. Metric: Agent Pass (%), higher is better.

Metric: Agent Pass (%). Source: tracelite.ai. Saturation forecast: Rough model projection: around 2026. 9 models tracked.

Top models

#ModelScoreOverall rank
1Claude Opus 5 (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)49.5
2Gemini 3.7 Flash (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)44.2
3GLM-5.3 (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)44.2
4Grok 4.6 (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)43.2
5Kimi K3 (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)43.2
6DeepSeek V4 Pro (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)37.9
7GPT-5.6 Sol (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)36.8
8Qwen 3.8 Max (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)32.6
9Hy3 (Benchmark Sweep Run: source=Trial-by-Trace; harness=DeepSeek Harness; snapshot-and-effort=unreported)28.4

Interactive version: theaggregate.ai/benchmark?slug=trial-by-trace-deepseek-harness-agent-pass · How It Works · Data refreshed daily, snapshot 2026-11-06.