Sign InOpen Brain
arXivPaperNeeds Review

An Empirical Study of Harness Design for Coding Agents

Harness components pay off differently by model and budget. Elide before summarizing, use planning selectively, and avoid elaborate tools when the model is already strong with bash.

arXiv · Sep 17, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Researchers tested **176 matched settings** across **four models**, varying planning, action space, and context management on SWE-Bench Verified and Terminal-Bench 2.1. Context handling mattered more as budgets tightened, chiefly by preventing overflow.

Practical Implication

Stage rule-based elision before LLM summarization. Use planning as an accuracy scaffold for weaker models and a cost control for stronger ones; offer predefined tools when bash skill is weak, but consider bash-only operation for capable models on CLI-heavy work.

Agent-Ready Context
Researchers tested **176 matched settings** across **four models**, varying planning, action space, and context management on SWE-Bench Verified and Terminal-Bench 2.1. Context handling mattered more as budgets tightened, chiefly by preventing overflow.

Stage rule-based elision before LLM summarization. Use planning as an accuracy scaffold for weaker models and a cost control for stronger ones; offer predefined tools when bash skill is weak, but consider bash-only operation for capable models on CLI-heavy work.

Recoverable elision added machinery without an accuracy gain because models rarely used recovery. The evidence spans **five context strategies** and **four window budgets**, but the abstract provides no effect sizes and covers only two benchmarks.
Connected Context · Feed7 Judgment

This supplies comparative evidence for choices that prior harness guidance largely framed as design principles. It narrows coding-agent optimization to model- and budget-dependent decisions: prioritize overflow prevention, begin with rule-based elision, use planning differently by model strength, and avoid recovery machinery whose availability does not translate into use.

Harness Engineering: Building the Production Cage for Powerful Domain Agents — Mike Chambers, AWSThe candidate defines the broad production-harness boundary; this study adds matched experimental evidence about which planning, tool, and context choices within that boundary matter for coding benchmarks.Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMindThe candidate argues that capable general tools can replace custom orchestration; this Signal qualifies that simplification because bash-only versus predefined tools depends on model capability and workload.Multimodal Collaborative Agents for Next-Gen Commerce — Nidhi Kaushik Vyas, Google DeepMindBoth treat context handling as an explicit harness policy, but this Signal isolates overflow prevention under constrained windows while the candidate focuses on eliciting missing user constraints.From coding to Knowledge work agents — Karan Vaidya, ComposioThe candidate extends harness controls to knowledge work; this Signal provides coding-specific empirical support that harness configuration materially affects outcomes, without establishing that the same rankings transfer across domains.
Context Map
agentcoding#harness-engineering#context-engineering#tool-use
Uncertainty
Recoverable elision added machinery without an accuracy gain because models rarely used recovery. The evidence spans **five context strategies** and **four window budgets**, but the abstract provides no effect sizes and covers only two benchmarks.