---
title: "Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite"
authors:
  - "Mohsen Arjmandi"
year: 2026
venue: "arXiv"
arxiv: "2609.11987"
doi: null
source: "https://arxiv.org/abs/2609.11987"
topics:
  - "agent-failure-localization"
  - "agentic-llm-serving"
seed_rank: 1822
seed_batch: "frontier-2026-09-14"
reviewed: "2026-09-14"
pool: "agents"
relevance_score: 10
lineage: runtime-harness-evolution
cites:
  - title: "Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite"
    url: "https://arxiv.org/abs/2609.11987"
    year: 2026
    arxiv: "2609.11987"
    doi: null
see:
  - "1798-ecdysis-efficient-and-effective-training-of-runtime-harnesses-for-llm"
  - "1536-harness-engineering-anatomy-architecture-and-evolution-of-coding-agent"
  - "1050-deepseek-harness-everything-is-a-plugin"
  - "1768-co-evolving-harnesses-and-models-on-policy-correction-helps-weaker"
  - "1573-where-does-harness-optimization-value-live-localized-gains-and-the"
---

# Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite

## One-sentence takeaway

Paired same-model contrasts on a private contamination-controlled suite show the coding-agent harness — not just the base LLM — moves a large share of solve rate.

## Why it matters here

Broadside’s standing model-vs-harness cut (1050 / 1536 / 1573 / 1768 / 1798) finally gets a clean measurement protocol: same model, swapped harness, private post-cutoff tasks. Anoptic tool buses and GRID COMMAND order harnesses should copy the paired-contrast discipline before claiming “model upgrade” wins.

## Key ideas

- Defines the agentic coding system as model + harness (tools, prompts, control flow) and measures the harness effect in isolation.
- Runs paired same-model contrasts on a private suite of 256 repository and post-cutoff contest tasks — contamination-controlled.
- Shows vendor-native harness pairings are not automatically best; harness choice moves solve rate under a fixed LLM.
- Gives Broadside a template for attributing failures to harness vs weights instead of blending them in leaderboard deltas.

## Caveats

Private suite limits external replication; abstract does not publish the full harness inventory. Do not remint Ecdysis 1798, Harness Engineering 1536, or DeepSeek Harness 1050 — this is the measurement companion, not another harness trainer.

## Links

- arXiv: [2609.11987](https://arxiv.org/abs/2609.11987)
- PDF: https://arxiv.org/pdf/2609.11987