A coding agent's harness shifts its cost and accuracy with the model and budget

TL;DR

• A 17 September arXiv paper varies planning, the tool set and context management across 176 settings on four models.
• Rule-based trimming before LLM summarisation was the most efficient; planning helped weak models and saved money on strong ones; bash-only cut cost for bash-capable models.
• The authors call the results conditional on their implementations, not a universal best harness.