cs.AINov 16, 2026

Harness Leverage Across Model Size: A Ladder of Five Local Agent Harnesses from 0.75B to 4.7B

Authors: Hao Wang, Ting Huang

Organizations: School of Mathematics and Physics, University of Science and Technology Beijing, Beijing, China · Honor Device Co., Ltd., Beijing, China

Abstract

A harness turns a model into a system that completes tasks: it builds the prompt, exposes tools and decides when work is done. Harness comparisons hold the model fixed. Model comparisons hold the harness fixed. We vary both: five shipped harnesses, 18 artifact-producing tasks, four locally deployable open-weight models from 0.75B to 4.7B and a 35.5B anchor, with one deterministic checker scoring all 360 ladder cells. The completion-controlling harness, Mingbird, has the highest mean at every size. Of four pre-registered paired contrasts, two survive Holm correction, at 0.75B and 4.7B. Thirteen of sixteen paired intervals exclude zero under the deployment reading, with leads from 0.075 over the strongest rival to 0.819 over the weakest. The completion gate is not the carrier: disabling all six executable checks moves the mean by +0.080, +0.117, +0.003 and -0.004, every interval including zero, and only at 1.9B does disabling it create empty runs. Removing one repeat-blocking rule alone costs 0.411 at 0.75B and leaves eight of eighteen tasks with no artifact. The harness effect is a progress effect: it decides whether a run leaves anything on disk. Of 357 usable cells, 127 leave none of their declared artifacts and 85 of those exit normally, so a scorer that reads exit status credits work that was never written. Whether the effect grows or shrinks with size, four rungs cannot settle: a floor arm produces a rise, a ceiling a decline, and a cross-family rung a peak.

Figures & tables

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 1, 2026

Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks

Small open-weight language models now run on a laptop, but they fail at multi-step work in ways a large model does not: they stop with the files missing, repeat one action until the budget ends, and report success without checking. At the one-to-nine-billion-parameter tier the harness supplies almost all of the procedural control that a large model supplies for itself.We present Mingbird, an open-source harness for that tier on a consumer graphics card: 12k lines of Python, a 628-test suite, and eighteen mechanisms for context economy, completion, time budget and the host boundary, plus a crawl guard outside those four groups. Its design point is completion control: the harness runs executable checks on the working directory and refuses a claim of completion when they fail.Five evaluations accompany it. A configuration-surface audit finds four defects that the 625-test suite of v3.1.1 missed, each because a test checked a property adjacent to the one that broke. A ten-task probe suite gives each mechanism a reason to fire and reports where the change reaches the artifact, including one mechanism that cuts the signal it protects. Disabling five of six completion-time checks costs 0.152 of the artifact score on eighteen held-out tasks, three replications per cell (95% CI [+0.064, +0.246], Holm p = 0.014). On a 288-cell matrix the harness reaches 0.886 against 0.631, 0.479 and 0.405 for three open scaffolds, and on τ 2 -bench it ranks first or tied first in all three domains against the arms evaluated.Tasks, scorer, source and every per-cell result are public.
Oct 3, 2026

What Does a Harness Buy? Tokens, Mostly

A coding agent is a language model inside a harness: the system prompt, the tools, and the context management that let a chat model work in a repository. Most of the engineering on these agents goes into the harness, and vendors and leaderboards credit it with score gains. Earlier studies have fixed the model and swapped the harness, but have not measured whether the swap changes the pass rate more than running the same harness again. We run seven models through three production harnesses, Claude Code, mini-SWE-agent, and OpenCode, on SWE-bench Verified and two benchmarks of other kinds, and rerun them unchanged. No harness leads consistently across models. Where the data can bound the difference between two harnesses, the bound is 2 to 5 points, and a harness swap changes about as many task outcomes as a rerun. The exceptions are harness effects in particular pairings of model and harness, whose direction changes with the model: OpenCode trails by 8 to 9 points on one model and leads by 14 on another. Cost differences between harnesses also depend on the model, from under 10% to 3 to 4 times. Which harness is better therefore has no answer apart from the model, and a harness difference should be reported next to a rerun.
May 27, 2026

Harness-Bench: Evaluating Agent Harnesses across Models and Realistic Workflows

LLM agents are increasingly deployed as executable systems that use tools, modify workspaces, and produce concrete artifacts. In such workflows, performance depends not only on the base model, but also on the harness: the system layer that manages context, tools, state, constraints, permissions, tracing, and recovery. However, existing benchmarks typically abstract away execution, compare complete agent systems, or hold the harness fixed, making execution-layer variation difficult to study. We introduce Harness-Bench, a diagnostic benchmark for evaluating configuration-level harness effects in realistic agent workflows. Harness-Bench evaluates representative harness configurations across multiple model backends under shared task environments, budgets, and evaluation protocols, while preserving each harness's native execution behavior. The benchmark contains 106 sandboxed offline tasks constructed from practical agent-use patterns and manually reviewed for realism, solvability, oracle-checkability, and integrity. Each run records final artifacts, execution traces, usage statistics, and validator outputs, enabling analysis beyond final completion. Across 5,194 execution trajectories, we observe substantial variation in completion, process quality, efficiency, and failure behavior across model-harness pairings. These results suggest that agent capability should be reported at the model-harness configuration level rather than attributed to the base model alone. Our analysis further identifies recurring execution-alignment failures, where plausible reasoning becomes decoupled from tool feedback, workspace state, evidence, or verifiable output contracts. Harness-Bench provides a reproducible foundation for diagnosing and improving reliable, efficient, and auditable agent execution stacks.