Harness Leaderboard: 28,849 trials show coding-agent harness choice can swing open-model scores by tens of points on Terminal-Bench
- Why it belongs in Context & Harness Attacks: the wrapper around the model — tools, prompts, compaction, permissions — is a first-class security and reliability surface, not a detail.
- Method highlights: three attempts with CIs, infra failures excluded, DeepSWE offline since Oct 23, cost/time metered outside the harness, secrets redacted.
- Companion: Harness Lens documents what each harness actually does, with sources.
Sources: Hugging Face Space: joelniklaus/harness-leaderboard · Harness Leaderboard README · Joël Niklaus on X (Nov 5)