---
title: "Claim Ledger: Same Model, Different Product: The Case for Harness Engineering"
description: "Hold the model fixed and the surrounding harness can still move task success, cost and reliability. Here are the layers that do it."
author: "Harry Floyd"
publication: "The Durability Curve"
canonical: "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/"
essay: "https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/"
substack: "https://harryfloyd.substack.com/p/harness-engineering-same-model-different-product"
published: "2026-05-30"
ledger_date: "2026-10-03"
law: "Law IV"
claims: 14
removed_before_publication: 0
---

# Claim Ledger: Same Model, Different Product: The Case for Harness Engineering

*Hold the model fixed and the surrounding harness can still move task success, cost and reliability. Here are the layers that do it.*

Essay: https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/  
Ledger (canonical, cite this): https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/  
Published 2026-05-30 · last verified 2026-10-03

## The evidence, row by row

Labels: VERIFIED = checked against the primary source itself by a checker who did not draft the piece, usually with the exact words, where they sit and the date read · EXECUTED = a run or observation done for the piece (code, a query, a count, or a check made live in an app); the claim is what it returned; run records are not published · CHECKED = checked against its source by a separate checker in the older ledger format, which recorded no quote, locator or date · REPORTED = not verified word for word against a primary source (secondary, not openable in full, or supported only in part); the essay words it accordingly · STRUCK = drafted, checked and removed before publication · EXCLUDED = considered and deliberately left out.

### HE1. In the Meta-Harness paper's TerminalBench-2 table, Claude Opus 4.6 has pass rates of 58.0 under Claude Code (an official-leaderboard result the paper reports) and 76.4 under the Meta-Harness harness. Scope: Meta-Harness searched and was finally evaluated on the same 89-task TerminalBench-2 benchmark, and the paper describes the resulting harness as specialised to the TerminalBench-2 regime.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE1
- Quote: search and final evaluation on the same 89-task benchmark … specialized to the TerminalBench-2 regime … Claude Opus 4.6 … Claude Code × 58.0 … Meta-Harness ✓ 76.4
- Source: Lee, Nair, Zhang, Lee, Khattab and Finn, Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv:2603.28052v1, p. 9, §4.3 Evaluating Agentic Coding Harnesses on TerminalBench-2 (setup paragraph) · Table 7, Claude Opus 4.6 block; caption: results of others are from the official leaderboard
- URL: https://arxiv.org/pdf/2603.28052

### HE2. In the same table, Claude Haiku 4.5 has pass rates of 27.5 under Claude Code (an official-leaderboard result the paper reports) and 37.6 under the Meta-Harness harness. Scope: Meta-Harness searched and was finally evaluated on the same 89-task TerminalBench-2 benchmark, and the paper describes the resulting harness as specialised to the TerminalBench-2 regime.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE2
- Quote: search and final evaluation on the same 89-task benchmark … specialized to the TerminalBench-2 regime … Claude Haiku 4.5 … Claude Code × 27.5 … Meta-Harness ✓ 37.6
- Source: Lee, Nair, Zhang, Lee, Khattab and Finn, Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv:2603.28052v1, p. 9, §4.3 (setup paragraph) · Table 7, Claude Haiku 4.5 block; also Figure 1 (right), p. 1
- URL: https://arxiv.org/pdf/2603.28052

### HE3. LangChain reports improving its coding agent deepagents-cli by 13.7 points on Terminal Bench 2.0, from 52.8 to 66.5, by changing only the harness while keeping the model fixed at GPT-5.2-Codex.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE3
- Quote: improve deepagents-cli … 13.7 points from 52.8 to 66.5 on Terminal Bench 2.0. We only tweaked the harness and kept the model fixed, gpt-5.2-codex
- Source: LangChain (Vivek Trivedy), Improving Deep Agents with harness engineering, 17 February 2026, opening section, paragraph beginning "We used a simple recipe"
- URL: https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering

### HE4. Claude Code's command reference documents /compact as a command that frees up context by summarising the conversation so far.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE4
- Quote: /compact [instructions] … Free up context by summarizing the conversation so far.
- Source: Claude Code Docs, Commands, command table, row "/compact [instructions]"
- URL: https://code.claude.com/docs/en/commands

### HE5. Claude Code's environment-variable reference documents CLAUDE_CODE_MAX_RETRIES, which overrides the number of times Claude Code retries failed API requests.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE5
- Quote: CLAUDE_CODE_MAX_RETRIES … Override the number of times to retry failed API requests
- Source: Claude Code Docs, Environment variables, Variables list, row "CLAUDE_CODE_MAX_RETRIES"
- URL: https://code.claude.com/docs/en/env-vars

### HE6. Claude Code's features overview describes subagents as isolated workers with their own context, running separately from the main conversation.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE6
- Quote: Subagents … are isolated workers that run separately from your main conversation … Isolated worker with its own context
- Source: Claude Code Docs, Extend Claude Code (features overview), Compare similar features, tab "Skill vs Subagent", row "What it is"
- URL: https://code.claude.com/docs/en/features-overview

### HE7. Manus warns, from its own experience building an agent rather than from a measured experiment, that when users plug hundreds of tools into an agent's action space, the model is more likely to select the wrong action or take an inefficient path.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE7
- Quote: hundreds of mysterious tools … the model is more likely to select the wrong action or take an inefficient path
- Source: Yichao 'Peak' Ji, Manus, Context Engineering for AI Agents: Lessons from Building Manus, 18 July 2025, section "Mask, Don't Remove", first paragraph
- URL: https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus

### HE8. Manus deliberately leaves failed actions and errors in the agent's context, arguing that erasing failure removes the evidence the model needs to adapt.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE8
- Quote: Erasing failure removes evidence. And without evidence, the model can't adapt. … leave the wrong turns in the context
- Source: Yichao 'Peak' Ji, Manus, Context Engineering for AI Agents: Lessons from Building Manus, 18 July 2025, section "Keep the Wrong Stuff In", second and third paragraphs
- URL: https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus

### HE9. In Can Bölük's edit-format benchmark, the hashline edit format beat the patch format in 14 of 16 models.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE9
- Quote: hashline beats patch in 14/16 models
- Source: Can Bölük, Stencil, We improved 15 LLMs at coding in one afternoon. Only the harness changed., 12 February 2026, benchmark chart caption (3 runs × 180 tasks per model)
- URL: https://stencil.so/blog/the-harness-problem

### HE10. In the same benchmark, Grok Code Fast 1 went from 6.7% with the patch edit format to 68.3% with hashline.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE10
- Quote: Grok Code Fast 1 went from 6.7% to 68.3%, a tenfold improvement … patch 6.7% … hashline 68.3%
- Source: Can Bölük, Stencil, We improved 15 LLMs at coding in one afternoon. Only the harness changed., 12 February 2026, results paragraph beginning "Sixteen models, three edit tools" · chart row 01 point labels
- URL: https://stencil.so/blog/the-harness-problem

### HE11. Menlo Ventures estimated in December 2025 that Anthropic held 54% of the enterprise coding market, up from 42% six months earlier, driven in large part by Claude Code. The figure comes from the report's LLM market-share analysis, which estimates model-provider share from production API usage.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE11
- Quote: coding market … an estimated 54% market share … up from 42% just six months ago, driven in large part by the popularity of Claude Code
- Source: Menlo Ventures, 2025: The State of Generative AI in the Enterprise, 9 December 2025, section "LLM Market Share: Anthropic Extends Its Lead in the Enterprise", paragraph after the share charts · methodology note "LLM Market Share"
- URL: https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/

### HE12. Anthropic said in February 2026 that Claude Code's run-rate revenue had grown to over $2.5 billion.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE12
- Quote: Today, Claude Code’s run-rate revenue has grown to over $2.5 billion
- Source: Anthropic, Series G funding announcement, 12 February 2026, paragraph beginning "Claude Code represents a new era of agentic coding"
- URL: https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation

### HE13. DeepSeek publishes DeepSeek Harness (dsh) from its deepseek-ai GitHub organisation, and the README describes it as an open-source agent harness developed by DeepSeek AI.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE13
- Quote: DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.
- Source: DeepSeek AI, deepseek-ai/deepseek-harness README, GitHub, README opening paragraph · DeepSeek's own site, deepseek.com/harness, links the repository
- URL: https://github.com/deepseek-ai/deepseek-harness

### HE14. The DeepSeek Harness README labels the project a developer preview.

- Status: VERIFIED (read 2026-10-03)
- Row link: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE14
- Quote: DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.
- Source: DeepSeek AI, deepseek-ai/deepseek-harness README, GitHub, README section "Developer preview"
- URL: https://github.com/deepseek-ai/deepseek-harness

## Cite

- A claim: name the original source (and its locator) first, then the row it was checked in, e.g. "<source>, <locator>. Checked in: Claim Ledger, "Same Model, Different Product: The Case for Harness Engineering", The Durability Curve, row <id>, https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-<id>."
- The essay: Floyd, Harry (2026). Same Model, Different Product: The Case for Harness Engineering. The Durability Curve. https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/
- This ledger: Floyd, Harry (2026). Claim Ledger: Same Model, Different Product: The Case for Harness Engineering. The Durability Curve. https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/
- Downloads: https://durabilitycurve.com/claims/harness-engineering-same-model-different-product.csv · https://durabilitycurve.com/claims/harness-engineering-same-model-different-product.json

Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Reuse it, quote it or train on it, with credit to The Durability Curve and a link. Open to AI (robots.txt: search=yes, ai-input=yes, ai-train=yes).