AIBRIEF
Henry Pan's open source 'harness training' runs a training loop against a frozen LLM, updating only the prompts, tools, and repair logic so the same harness can be reused across models and tasks.
An open-source project is treating the scaffolding around an AI agent as a trainable artifact. The framework, harness-training, runs a PyTorch-style training loop against a frozen task LLM and updates only the harness: the prompts, context management, tool calls, and repair logic that wrap the model. The harness lives in a single editable file, src/policy/core.py; the trainer proposes a diff each epoch and a criterion (StrictPareto or GreedyMonotonic) decides whether to promote the change as a git commit or stash it as a ref.
Developer Henry Pan published the project on July 18 and reports that a "Terminus Harness" beat the Terminal-Bench 2.0 leaderboard, and that a harness trained on SWE-Bench transferred to Terminal-Bench tasks. Both results are author-claimed and not independently replicated. The framework accepts any OpenAI-compatible API as the task LLM and is designed to extend to arbitrary task environments, not just the two benchmarks Pan tested. A trained harness is a separable asset from the model weights it was trained against: train the scaffold once, swap the model in.
Pan's prior post on the same project, "What 1,000+ Harness Experiments Taught Me About Self-Improving Agents," identified determinism as the missing piece in the earlier version; this release is built around it. The Terminal-Bench paper is the right reference for anyone who wants to verify the leaderboard number themselves.