Reasoning effort, not tool access, buys first-try reliability in agentic code generation: an observational study

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Achint Mehta
• Venue: arXiv: cs.SE
• Published: 2026-07-02
• Collected: 2026-07-03
• URL: http://arxiv.org/abs/2607.02436v1
• PDF: https://arxiv.org/pdf/2607.02436v1
• DOI: UNKNOWN

Plain-Language Brief

A study tested the assumption that giving agentic coding assistants more capabilities leads to better software. The study found that investing in reasoning effort yielded better first-try reliability and fewer corrective prompts, while adding a testing tool increased cost without improving functional score or reliability.

What the Document Says

• The study involved 90 independent agent runs building the same application using one detailed specification.
• Each run was scored on a 14 criterion functional rubric and visual quality review.
• Different model generations, agent harnesses, reasoning effort levels, testing tools, and design oriented prompts were used.
• Frontier models clustered near the ceiling of the score range, while a low-cost local model scored lower.
• Container deployment was found to be a dominant defect, failing in 44% of runs and its failure rate shifting across model generations.
• The testing tool increased cost but did not improve functional score or reliability.
• Raising reasoning effort improved first-try perfect runs and cut corrective prompts.
• A design oriented prompt improved visual quality without lifting function.
• The study suggests that most first run failures come from weak reasoning, which can be addressed by a stronger model or more effort.

Key Concepts Mentioned

• agentic coding assistants
• reasoning effort
• testing tool
• design oriented prompts

Claims Made by the Document

• More capability does not necessarily yield better software.
• Investing in reasoning effort improves first-try reliability and reduces corrective prompts.
• Adding a testing tool increases cost without improving functional score or reliability.

Evidence Used

• 90 independent agent runs
• 14 criterion functional rubric
• visual quality review

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Agentic coding assistants are increasingly given extra capabilities, such as browser based testing tools and design oriented system prompts, on the assumption that more capability yields better software. This study tested that assumption directly. Ninety independent agent runs built the same application, a real time retrospective board, from one detailed specification, each scored on a fixed 14 criterion functional rubric (42 point maximum) and a visual quality review. The runs spanned several model generations, two agent harnesses, two reasoning effort levels, a testing tool, and two design oriented prompts. Capability tier dominated: frontier models clustered near the ceiling while a low cost local model fell to 24 to 37 points. A criterion level analysis revealed what run totals conceal. Container deployment was the dominant defect, failing first try in 44 percent of runs, with its failure rate shifting sharply across model generations while mean totals moved less than a point. The testing tool raised cost by 42 to 68 percent without improving functional score or reliability, even on interface visible criteria. Raising reasoning effort from High to xHigh lifted first try perfect runs from 28 percent to 89 percent and cut corrective prompts about five fold, for 9 to 29 percent more cost. A design oriented prompt raised visual quality, 4.5 versus 3.0 on a 5 point scale, without lifting function, and a one paragraph paraphrase of its directive reproduced the entire lift. The practical lesson is to match the fix to the failure: most first run failures came from weak reasoning, which a stronger model or more effort prevents, not from visible flaws a checking tool would catch.