A 176-configuration harness ablation says rule-based elision before LLM summarization wins, and recoverable deletion barely helps.

arXiv 2609.20804 holds the execution loop fixed and varies planning, action space and context management independently across four models. Context management matters most under tight token budgets, the winning recipe is rule-based elision applied before any LLM summarization, and making the elided text recoverable adds almost nothing. Planning acts as accuracy scaffolding for weaker models and only a cost reduction for stronger ones. Bash-only interfaces beat predefined tools for bash-proficient models. These are configuration rules you can apply, not a general "better harness" claim.