Coordination benchmark: multi-model vs single-model on 49 coding tasks (open receipts)

[WillNigri]

Open coordination benchmark (2026-07)

We published receipt-backed results on multi-model coordination vs single-model routing for coding.

Headline: higher floor, lower ceiling. On 49 contamination-audited LiveCodeBench tasks, no coordination recipe beat the best single model. Coordination still lifted weak models (~+9 problems). Cascades were mostly limited by a weak verifier gate, not the strong closer.

Read

• Human summary: https://agentictool.ai/posts/does-multi-model-coordination-beat-a-single-model
• Full draft paper (HTML): https://agentictool.ai/posts/coordination-benchmark-paper.html
• Markdown / PDF: https://agentictool.ai/posts/coordination-benchmark-paper.md · https://agentictool.ai/posts/coordination-benchmark-paper.pdf
• In-repo map: https://github.com/WillNigri/Agentic-Tool-Optimization/blob/main/docs/research/README.md
• Open-box re-run path: ato bench run + cascade recipes — see README “open-box router”

What we want from this thread

• Critiques of method (contamination, n=49, public-test gates, war-room shape)
• Replications with your model set / keys
• “I already multi-model review” war stories — especially Claude + Codex / Gemini

Soft CTA

If you use two or more coding runtimes and want a 15-minute white-glove pass on your PR (ato review --consensus style), say so here or reach out — looking for reality checks, not vanity installs.

Status: draft technical report (revisions 2026-07-10 / 2026-07-11). Directional at this n; the suite and receipts are the product.