SWEBench
Harbor: pendingThis adapter converts SWEBench Verified benchmark instances into Harbor-compatible tasks , enabling evaluation of reasoning agents within isolated, reproducible software environments.
- Samples
- 500
- Released
- 2025
- Scores
- 0
Doc2Harness
Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents
A two-stage benchmark evaluating LLMs' ability to engineer coding agent harnesses from documentation and interface skeletons, scored via 328 test cases.
How to run
ltb · doc2harness
$ npm i -g lettokenburn-cli$ ltb bench info doc2harness
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
14 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-6 Astra | 53.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Claude Opus 5 | 46.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Qwen3.8 Max | 45.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.6 Sol | 43.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Grok-4.7 | 43.5 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GLM-5.3 | 42.1 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| MiMo V2.6 Flash | 41.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Step 5 Preview | 40.8 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 40.2 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| HY4 Preview | 38.9 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Kimi K3 | 38.3 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Claude Opus 4.8 | 38.0 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| GPT-5.5 | 34.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
| Doubao-Seed-2.1-Pro | 32.6 | N/A | Independent | medium | Oct 14, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Doc2Harness evaluates long-horizon harness engineering by having models build a candidate harness from a development document and contract skeleton in Stage 1. In Stage 2, the harness is driven by a mock model across isolated containers.
Scoring uses four observable channels (requests, files, exit codes, output) with positive controls to prevent vacuous passes. GPT-6 Astra achieved the highest pass rate of 53.05% across 328 test cases.