One harness trained on a small model lifts Terminal-Bench scores across four models
Megadragon9 · reddit · 2026-07-24
Training the harness on one small model lifts Terminal-Bench scores across four others
A developer reports that training an agent harness—not the base LLM itself—improved Terminal-Bench 2.0 scores for four different models, including DeepSeek v3.2, GPT-OSS 20B/120B, and MiniMax M2.5, and even beat the official Terminus 2 harness.
The setup freezes the model and optimizes the surrounding system:
- prompts
- tools
- context management
- repair loops
The harness lives in a single Python file, and the training loop proposes changes, then promotes them only if they meet a quantitative criterion.
The author trained the harness once using a small Qwen 3.6 35B A3B NVFP4 model hosted on rented 5090 GPUs. They then used that trained harness across different task environments, arguing that the gains transfer beyond the model used for training.
The accompanying framework is presented as a more general PyTorch-like training loop for agent harnesses, with OpenAI-compatible backends and support for Terminal-Bench or SWE-Bench, plus extension points for other environments. The post also says the biggest missing ingredient in the first version was determinism.
More from coding & agent
- How should an MCP app be designed for non-technical users? — sn0wquake · 2026-07-24
- Greg Kamradt highlights disposable URLs as a new primitive for agent workspaces — GregKamradt · 2026-07-24
- When should you use a dedicated memory framework for agents? — markotkid · 2026-07-24
- Sentry’s Seer agent answered a Slack question with 30 days of Notion MCP usage data — zeeg · 2026-07-24
- Greg Kamradt says vibe coding raises both the floor and the ceiling — GregKamradt · 2026-07-24
- This AI brief writes your actions back so tomorrow’s summary gets smarter — Spirited_Ad_3886 · 2026-07-24