English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Paper: Harness or Model? Isolating the Harness Effect in Agentic Coding

Forum topic · 小凯 · 2026-09-15

Summary

This arXiv paper (2609.11987) tests the common assumption that vendor-native agent harnesses outperform neutral harnesses in agentic coding. The authors ran paired same-model contrasts on a private, contamination-controlled suite of 256 repository and post-cutoff contest tasks: claude-opus-4-8 under claude-agent-sdk vs deepagents, and gpt-5.5 under the openai-codex SDK vs deepagents, with gemini-3.5-flash and deepseek-v3.2 as side cells. Of 800 planned runs, 792 were graded by an isolated oracle. Neither pairing showed an average advantage (Opus: -1.25 pp, CI [-10.0, +7.5]; GPT-5.5: +1.25 pp, CI [-4.4, +6.9]), though Opus results split sharply by task type: vendor harness lagged 9.0 pp on repository tasks but led 23.7 pp on contest tasks (p=0.003). Cost re-pricing suggests the neutral harness is 1.2-1.6x more expensive per solved task, though billing ranking remains unresolved due to missing usage records. Orchestrator, scoring oracle, and analysis code are open-sourced.

Paper Overview

  • Field: NLP
  • Author: Mohsen Arjmandi
  • Published: 2026-09-15
  • arXiv: 2609.11987
  • Abstract (English translation)

    An agentic coding system couples a language model to a harness: the tools, prompts and control flow that turn a chat model into an autonomous software engineer. Vendors ship harnesses tuned to their own models, and practitioners assume the vendor-native pairing solves more tasks.

    The authors measure that assumption with paired same-model contrasts on a private, contamination-controlled suite of 256 repository and post-cutoff contest tasks. The same 80 tasks ran under claude-agent-sdk and under deepagents on claude-opus-4-8, and under the openai-codex SDK and deepagents on gpt-5.5, with gemini-3.5-flash and deepseek-v3.2 as side cells. 792 of 800 planned runs were graded by an isolated oracle.

    Key findings

  • No average harness advantage: Neither contrast resolves an average advantage for either harness: -1.25 pp for Opus 4.8 (48.8% vs 50.0%, task bootstrap 95% CI [-10.0, +7.5]); +1.25 pp for GPT-5.5 (55.6% vs 54.4%, CI [-4.4, +6.9]).
  • Stratified reversal for Opus: The vendor harness trailed by 9.0 pp on 61 repository tasks but led by 23.7 pp on 19 contest tasks (label-permutation p=0.003). This split was chosen after seeing the data and awaits a designed replication.
  • Correctness vs completion diverge: Of 81 runs canceled at a clock cap, 22 had already produced test-passing patches.
  • Cost estimates: Re-priced from raw per-turn usage at frozen catalog prices, the neutral harness costs 1.3-1.6x more per solved task on Opus 4.8 and 1.2x more on GPT-5.5. These are observational estimates; 58 runs on an Anthropic account left no usage records, so the billing ranking remains unresolved (Opus ratio could shift between 0.7 and 2.3).
  • Revision note: This revision corrects an August 2026 manuscript whose cost figures were based on a usage-semantics defect in the authors' own telemetry (Section 5.1).
The orchestrator, scoring oracle, re-analysis code, and derived aggregate data are open-sourced; the tasks remain private.

--- *Auto-collected on 2026-09-15*

Tags

#agentic-coding#llm-benchmarks#arxiv#nlp#evaluation#software-engineering#harness

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634828