Harness comparisons
11The same task in different coding harnesses: the agent program around the model, such as Claude Code, Codex, Cursor or OpenCode. 8 of 11 keep the model fixed, so only the harness changes. The rest pair each harness with its own model, so harness and model effects are mixed.
Harnesses
| Harness | Maker | Posts |
|---|---|---|
| Claude Code | Anthropic | 5 |
| Cursor | Anysphere | 3 |
| OpenAI Codex | OpenAI | 3 |
| OpenCode | SST | 3 |
| DeepSeek HarnessNamed as in the posts. | DeepSeek | 2 |
| Grok Build | xAI | 2 |
| Gemini CLI | 1 | |
| Google Antigravity | 1 | |
| Kimi CLI | Moonshot AI | 1 |
| OMPNamed “omp” / “OMP agent framework” in the posts; the maker isn’t stated. | Not stated | 1 |
| PenguinHarnessOpen-source agent harness (github.com/Prism-Shadow/penguin-harness). No logo with a clear licence, so a monogram is shown. | Prism Shadow | 1 |
| PiNamed only as “Pi” in the post; most likely the open-source pi coding agent, which the post doesn’t confirm. | Not stated | 1 |
| Tencent WorkBuddyTencent's AI agent desktop workbench (the CodeBuddy family). The logo is Tencent's mark. | Tencent | 1 |