Harness Engineering for AI Agents

The central claim — the harness matters more than the model — and the pipeline architecture, experimental methodology, and model-role separation that put it on rigorous footing.

Planning, dual-search, saturation gating, persistent memory, and large-document context extraction — the research half of the pipeline that determines what the model sees.

The evaluate → revise → verify loop, the decimalized 5-dimension rubric, harness-side verification, failure taxonomy from 1,500 logged runs, and the 3-persona evaluation panel.

The skills hook system, multi-subtask orchestration, security layer (AST scanning, path sandbox, injection detection), cost and inference management, and structured observability.

Autonomous synthesis optimization (autoresearch), the full data pipeline from runs.jsonl to SFT and DPO datasets, QLoRA fine-tuning, and the literature review pipeline.