Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.04433

What Does a Harness Buy? Tokens, Mostly

Yangze Liu·Zhongyi Han
View original source ↗Read source document ↗Export metadata ↓

Abstract

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. Production harnesses ship releases daily, 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. We run five models through three production harnesses, Claude Code, mini-SWE-agent, and OpenCode, on SWE-bench Verified, and rerun the same configurations to calibrate how much a score moves when nothing changes but the run. On 447 tasks and the two models we ran there, Claude Code and mini-SWE-agent, the heaviest and the lightest harness, are equivalent within five points. On a 45-task hard subset and five models, swapping the harness flips as many tasks as rerunning the same harness, 13% in both cases, and the tasks a harness wins in one run are not the tasks it wins in the next. The one harness effect that clears the noise is a loss, not a gain: OpenCode trails by up to 9 points on the large pool, and on one model half of that gap sits in runs its output cap cut short. What the harness does decide is the bill. With the same model, the same tasks, and one price list, cost per task differs by up to 3x across harnesses. The gap is set at the first call, by the preamble of system prompt and tool schemas each harness sends with every step, and scaled by the number of steps; per-step growth and per-call tool output differ far less. The provider's price for cached input scales the bill and does not reorder it. The rerun data also give the resolution a harness comparison needs: at the discordance we observe, 45 tasks catch a 13-point gap only half the time and no gap with 80% power, and 447 tasks resolve 5 points, still coarser than the gains many harness changes claim.

Explore related subjects

cs.AIcs.LGcs.SE

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yangze Liu, Zhongyi Han. 2026-10-03. What Does a Harness Buy? Tokens, Mostly. https://arxiv.org/abs/2610.04433

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

PaTAS: A Framework for Trust Propagation in Neural Networks Using Subjective Logic

Trustworthiness has become a key requirement for deploying artificial intelligence in safety-critical applications, yet conventional metrics such as accuracy fail to capture uncertainty or the reliability of predictions, particularly under adversarial or degraded conditions. This paper introduces the Parallel Trust Assessment System (PaTAS), a framework for modeling and propagating trust in neural networks using Subjective Logic (SL). PaTAS operates in parallel with standard neural computation through Trust Nodes and Trust Functions that propagate input, parameter, and activation trust across the network, refining parameter trust during training (Parameter Trust Update) and returning a per-inference trust opinion, quantifying belief, disbelief, and uncertainty, on each prediction (Inference-Path Trust Assessment). Experiments on real-world and adversarial datasets show that these estimates are interpretable and stable and complement accuracy: PaTAS flags predictions that remain accurate yet rely on parameters learned from corrupted data, detects adversarially patched inputs without knowledge of the trigger, and audits training-label corruption without any test data. Compared with output-only uncertainty baselines and dedicated input- and feature-space detectors, out-of-distribution and corruption detection is carried by the conformity-derived input opinions PaTAS consumes, within a regime that a per-dataset diagnostic identifies in advance; trust propagation adds model-side signals these detectors lack, and the composed trust degrades gracefully where any single signal fails. The mechanism extends to convolutional networks. PaTAS thereby provides a foundation for transparent, quantifiable trust reasoning across the AI lifecycle.

cs.AI↗
arXiv2026

High-Precision Estimation of the State-Space Complexity of Shogi via the Monte Carlo Method

Determining the state-space complexity of Shogi (Japanese Chess) has been a challenging problem, with previous combinatorial estimates leaving a gap of five orders of magnitude ($10^{64}$ to $10^{69}$). This gap arises from the difficulty of distinguishing positions reachable from the initial position among the vast number of board configurations. In this paper, we present a high-precision statistical estimate of the number of reachable Shogi positions. Here, a position consists of the side to move, the board configuration, and the pieces in hand, without move history, and positions related by exchanging the two players or by horizontal reflection of the board are counted once. Our method combines Monte Carlo sampling with a reachability test that performs reverse search toward the set of ``King-King only'' (KK) positions rather than toward the single initial position. Since every KK position and the initial position are mutually reachable, reaching any KK position proves reachability, and removing pieces from the board is a much simpler objective than reconstructing the initial position. Based on a sample of five billion positions, we estimate the number of such positions in Shogi to be $6.55 \times 10^{68}$; both ends of the approximate $3σ$ confidence interval round to this value at three significant digits. We also applied this method to Mini Shogi, obtaining approximately $2.38 \times 10^{18}$.

cs.AI↗
arXiv2026

DERM-3R: A Resource-Efficient Multimodal Agents Framework for Dermatologic Diagnosis and Treatment in Real-World Clinical Settings

Skin diseases impose a substantial and growing global health burden. Modern dermatologic therapies control acute manifestations rapidly, but single-target treatment paradigms, recurrent disease courses, and neglected systemic comorbidities limit their long-term effectiveness. Traditional Chinese medicine (TCM) offers a complementary, holistic approach through syndrome differentiation and individualized treatment, yet its clinical practice is hindered by non-standardized knowledge, incomplete multimodal records, and the difficulty of scaling expert-driven reasoning. We propose DERM-3R, a resource-efficient multimodal agents framework that models TCM dermatologic diagnosis and treatment under limited data and computational resources. We decompose and redefine real-world dermatologic decision-making into three essential issues: fine-grained lesion recognition, multi-view lesion representation with specialist-level pathogenesis modeling, and holistic clinical reasoning for syndrome differentiation and treatment planning. DERM-3R comprises three collaborative agents, DERM-Rec, DERM-Rep, and DERM-Reason, each addressing one of these issues. Built on a lightweight multimodal large language model and fine-tuned by partial-parameter finetuning on 103 real-world TCM psoriasis cases, DERM-3R performs strongly across multiple dermatologic reasoning tasks. Automatic metrics, LLM-as-a-Judge assessments, and human doctors show that, despite extremely limited training data and parameter updates, DERM-3R matches or even surpasses hundred-billion-parameter general-purpose multimodal models such as GPT-5.1 and Gemini-3-Flash. Our results indicate that structured, domain-aware multi-agent modeling is an effective alternative to brute-force scaling for complex clinical tasks, offering a practical and scalable paradigm for multimodal AI in dermatology and integrative medicine.

cs.AI↗