From Code Generation to Execution-Guided Debugging: The Next Frontier in AI-Assisted Software Engineering

Read full text Summarise with VIRAT
Authors: Rajeshkumar Venugopal
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

AI has solved the easy half of software engineering. Code generation — the production of syntactically correct, functionally plausible code from natural language — is now a commodity capability. What remains unsolved is the hard half: understanding why code fails, under what conditions, and how to stop it from failing again. This paper argues that the field has been measuring progress on the wrong axis. The cost of software defects is not proportional to the cost of writing code; it is proportional to how late in the lifecycle those defects are discovered. The dominant cost in software engineering is not production — it is escape. A defect that reaches production costs between 40 and 100 times more to resolve than a defect caught in design. AI systems that increase code throughput without improving defect detection rates do not reduce this cost. They accelerate it. We propose execution-guided single-step debugging as the correct next objective: AI systems that ingest runtime logs, distributed execution traces, and infrastructure state to reconstruct the causal chain of a failure, identify the minimal invariant violation, and produce a bounded corrective action. We define this capability formally, propose multithreaded concurrency bugs as the appropriate evaluation benchmark, and argue that the field's central challenge is now a data problem — one that must begin not with repository scraping but with structured extraction from version-controlled bug histories.

Fields of Study & Subjects

Software Testing and Debugging Techniques Software Engineering Research Software Engineering Techniques and Practices

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
23 March 2026
Language
EN