A coding agent is a language model wrapped in a harness: the system prompt, the tool set, and the context management that turn a chat model into something that can work inside a repository. Vendors advertise pass-rate gains from harness changes, and leaderboards mix harnesses freely. What is rarely measured is how much the harness itself moves the score when the model is held fixed.
Researchers at Shandong University—Yangze Liu and Zhongyi Han—ask that question in *What Does a Harness Buy? Tokens, Mostly.* (arXiv:2610.04433). They pin three production harnesses—Claude Code, mini-SWE-agent, and OpenCode—and run the same models through all three on SWE-bench Verified, with reruns so a harness gap can be read next to a rerun gap. Models include Qwen3.6-35B-A3B, Qwen3.8-27B, DeepSeek-V4-Flash, GLM-5.3-Flash, and HY4-Preview, with Claude Opus 5 as an anchor inside Claude Code only.
Pass rate barely moves
On the 447-task pool, Claude Code and mini-SWE-agent—the heaviest and the lightest harnesses—are equivalent within five points on both Qwen models they report there. On the 45 hardest tasks, swapping the harness flips about as many tasks as rerunning the same harness (about 13% at the median), and the tasks a harness wins in one run are not the ones it wins in the next. In short: on this design, the harness is not buying a stable pass-rate win.
Where OpenCode trails
The gaps that clear the noise mostly have OpenCode on the losing side—by up to about nine points on the large pool for one Qwen model. The paper traces much of that to an output-token cap with no recovery: when a step is cut mid-reasoning, OpenCode does not re-prompt, and many of those trials end with nothing on disk. Claude Code hits similar caps but resumes. Scoring only tasks where OpenCode was not cut off roughly halves its gap. That is a concrete failure mode, not a blanket ranking that OpenCode is “worse everywhere.”
What the harness does buy: the bill
With the same model and the same tasks under one price list, cost per task differs by up to about 3× across the three harnesses. Little of that comes from tool returns—per call, the three hand the model about as much text. It comes from the fixed preamble each harness re-sends every step (system prompt and tool schemas: on the order of ~16k tokens in Claude Code, ~7k in OpenCode, and under a thousand in mini-SWE-agent), scaled by how many steps the model takes. Cache-hit pricing scales the bill and does not reorder the three harnesses in the authors’ checks.
How this sits next to install guides and multi-agent scale
Install steps for DeepSeek’s own harness are in DeepSeek Harness Install Guide: Windows, Linux, CLI, Plugins. Product-side Copilot harness defaults in VS Code are in VS Code 1.143 Makes the Copilot Harness the Default for New Sessions. The other axis—keeping one per-agent loop and changing how many workers share the job—is the pitch in Same Copilot Loop, More Agents: Agensh Claims Pass Rates Rise Toward 1,024 Workers.
Figures and cost multiples above are from the paper and its public trial table; we have not re-run the grid. For reference only. Feedback on similar harness swaps can go through the site-owner contact link in the footer.