Intel / article

Do Coding Agents Still Need a Debugger? Frontier-Model-Based Fault Localization at the Accuracy Ceiling

Source: arxiv.org
Author: Christoph Daxerer, Johann Glock, Christian Macho, Rudolf Ramler, Martin Pinzger, Josef Pichler
Date: Published Oct 28, 2026

Why it matters

If frontier agents already localize nearly all single-line faults from source alone, adding debugger tooling may cost more than it returns on this class of bugs. It helps decide where to spend tool-integration effort.

Key takeaways · AI-distilled

The authors report that without tools, agents found 98.2 to 99.9% of faults in the original programs and 96.7 to 99.9% after behavior-preserving rewrites, with most of the lost accuracy where the rewrites restructured loops and conditions.
Encouraging debugger use removed 63% and 61% of the remaining errors on rewritten programs for Luna 6 and Qwen 3.8 (+2.1 and +1.3 points), but cost 14.7 to 27.0 times the , per the abstract. Sol 6.1 stayed at 99.9%.
When the debugger was optional, agents stepped through code in only 0.2 to 6.2% of runs, yet Luna 6 and Qwen 3.8 still gained 1.5 and 1.3 points. The authors say agents mostly ran the program rather than stepping through it and kept most of the benefit.
The authors limit their conclusion to short, precisely specified programs: single-line faults seeded into LeetCodeDataset solutions. Whether source-only localization holds up on larger, loosely specified codebases is outside what this study measured.

Terms in this piece · Glossary

token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source arxiv.org