---
title: 'In the News: November 8, 2026 (Extra 3)'
description: 'A public leaderboard finds five coding-agent harnesses statistically tied on open models, with a 24-point gap between best and worst.'
canonical_url: 'https://darkfactory.dev/news/2026-11-08-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-08-extra-3.md'
collection: news
date_published: '2026-11-08T23:10:00-05:00'
date_modified: '2026-11-08T23:10:00-05:00'
---

# In the News: November 8, 2026 (Extra 3)


A new public leaderboard ran nine coding-agent harnesses against the same open-weight models and found the top five statistically tied. The gap that holds up is between that group and the bottom of the table.

## 1. A leaderboard puts five harnesses in a statistical tie on open models

**[Harness Leaderboard: Which coding-agent harness should you run your open model in?](https://huggingface.co/spaces/joelniklaus/harness-leaderboard)** · Hugging Face Space published by joelniklaus · version v0.1.1, surfaced on Hacker News November 8, 2026

The leaderboard runs nine harnesses on three open-weight models (GLM-5.3-Flash, DeepSeek-V4.1-Flash and Flint) across Terminal-Bench 2.1 and DeepSWE v1.1. Each harness gets the same model, provider and reasoning effort, three attempts per task, and every trajectory is published. On its macro average, mini-swe-agent leads at 51.4% (95% CI 48.5% to 54.2%) and OpenClaw is last at 27.3%, a 24-point range.

The page marks Kimi Code, Hermes Agent, Pi and Claude Code as tied with the leader, because the confidence interval of each score difference covers zero. Qwen Code (6.3 points behind), Codex (7.5), OpenCode (16.2) and OpenClaw (25.7) trail. On cost per solved trial, Pi is cheapest at $0.29, against $0.53 for mini-swe-agent and $0.52 for Claude Code.

The author also measured noise. Re-running the same harness flips 23.4% of tasks, and swapping to a different harness flips 29.1%. The page says harness choice adds 5.7 points of flips beyond rerun noise. Its method states that a trial that looks up a benchmark's published answers counts as a fail.

**Why it matters:** If you are choosing among the top five on these models, the page's own statistics say the ranking is not the reason to pick one. Cost per solve and reliability across attempts are better grounds. The results come from open-weight models, not frontier ones, and Claude Code and Codex were run against those models here, so they say little about how those harnesses perform with their vendors' own. The figures are the author's own and have not been replicated; we read the published page but did not open the trajectories.