1369-UC LLM-Based Evaluation and Feedback for Agentic Software Debugging

Presenter Information

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Large language models (LLMs) are increasingly being incorporated into software engineering workflows, including automated program repair. However, generating a plausible patch is not sufficient to establish that a repair is correct. The senior capstone project associated with this proposal develops and agentic software debugger and documenter in Python using Google ADK, with an Analyst Agent, a Fixer LoopAgent, deterministic validation through syntax checks, Ruff/Pylint, and Pytest when tests are available, and a Documentation Agent. My extension (honors portion) of the project, extends that system with an LLM-based Evaluation and Feedback Layer. Rather than replacing deterministic validation or adding another repair loop, the proposed layer will interpret evidence produced by each repair attempt, classify unresolved failure, assess possible regressions, identify root causes, and return structured feedback that can be consumed by the existing Fixer.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1369-UC LLM-Based Evaluation and Feedback for Agentic Software Debugging

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Large language models (LLMs) are increasingly being incorporated into software engineering workflows, including automated program repair. However, generating a plausible patch is not sufficient to establish that a repair is correct. The senior capstone project associated with this proposal develops and agentic software debugger and documenter in Python using Google ADK, with an Analyst Agent, a Fixer LoopAgent, deterministic validation through syntax checks, Ruff/Pylint, and Pytest when tests are available, and a Documentation Agent. My extension (honors portion) of the project, extends that system with an LLM-based Evaluation and Feedback Layer. Rather than replacing deterministic validation or adding another repair loop, the proposed layer will interpret evidence produced by each repair attempt, classify unresolved failure, assess possible regressions, identify root causes, and return structured feedback that can be consumed by the existing Fixer.