Research

I study how developers and AI coding agents resolve software issues, and I build techniques and tools that help them.

AI coding agents for UI bug repair

What agents need from a bug report, and how they repair bugs that users see on the screen.

  • Bug report content for coding agents

    We ran a controlled study of how the content of a bug report affects AI coding agents that repair Android app bugs. We used about 50 real bugs from over 20 apps, three agents (Codex, Claude, and Qwen), and more than a dozen report configurations, for over 7,500 repair attempts.

    Steps to reproduce had positive estimated effects for all three agents. Code localization reduced the median repair cost by over 30% for both Codex and Claude.

  • How agents repair UI bugs

    A companion study analyzed hundreds of repair attempts on the same bugs. Agent patches passed the tests in about three-quarters of attempts. About one in seven attempts matched the developer's fix. Nearly all other test-passing patches were valid alternative fixes, and very few overfit the tests.

  • Tools and data

    INCEPTOR is a browser-based tool for exploring the trajectories of program repair agents. The SE AI Agent Index is a dataset of software engineering practices for agentic AI systems.

Bug report quality and reproduction

Techniques that connect bug reports to the app's UI and execution, so reports are easier to reproduce and fix.

  • AstroBR

    AstroBR combines LLMs with dynamic app analysis to assess the steps to reproduce in bug reports. It outperformed the state of the art by 25.2% in S2R quality annotation and by 71.4% in missing-step identification.

  • BugScribe

    BugScribe grounds an LLM in the app's execution data to improve user-written bug reports. It improved the quality of the steps to reproduce by 44.2%–84.7% and raised an LLM agent's bug reproduction rate by 28.1%–57.7%.

  • Buggy UI localization

    We introduced the task of finding the buggy screens and UI components from a bug description, with a benchmark of 228 bug descriptions from 39 Android apps. Using the localized screens ranked the buggy code up to 12% more accurately.

  • Burt++

    Burt++ is a conversational bug reporting agent. It asks follow-up questions grounded in the app's GUI until each part of the report is complete. The paper is in preparation.

    In preparationCode

Issue resolution in practice

How developers resolve issues in real projects, and tools that support the hardest steps.

  • Firefox

    We analyzed 356 Firefox issue reports, built a model of the resolution process, and derived a catalog of 47 resolution patterns.

  • AI/ML systems

    We studied 100 issues and pull requests from four AI/ML systems (TensorFlow, scikit-learn, MLflow, and AutoGPT). Fixes change datasets, prompts, and model configurations, not only code.

  • Solution identification

    We compared traditional ML models, pre-trained language models, and LLMs in 68 configurations on 4,917 annotated comments. The best model, a fine-tuned LLM, found 83.8% of the solution-related comments.

  • Code change rationale

    We traced where rationale is documented across the artifacts of 63 commits and built ARGUS, an LLM-based approach that recovers it. ArgusVS brings this into the IDE.

  • Sprint

    Sprint is a GitHub assistant that finds similar issues, predicts issue severity, and suggests code files to change.