Agent Toolkit For Efficient Feature Creation Within
Existing Codebase
Name
matapayerro-anamp-meng-eecs-2026-thesis.pdf
Size
2.71 MB
Format
Adobe PDF
Checksum (MD5)
30feb4c242e9db8183220c9efb20c11a
Author(s)
Mata Payerro, Ana C.
Advisor(s)
Abelson, Hal
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
Large language model (LLM)-based agents have shown promise in automating software development tasks, but onboarding such agents into existing codebases remains challenging due to limited context windows and the structural complexity of real-world repositories. This thesis investigates whether feature-based representations of a codebase can improve an agent’s ability to locate and modify relevant code when implementing new features. To explore this question, an agent system was developed that represents a code repository as a graph of files, classes, and functions, which the agent navigates using feature-aware tooling. Two feature encoding strategies were evaluated. The first, a keyword-based feature encoding, used repository summaries to generate features and link them to code through keyword matching. The second, a cluster-based feature encoding, grouped code components using latent semantic analysis and constructed feature relationships based on semantic similarity. Both approaches were evaluated on feature implementation tasks drawn from the SWE-Bench dataset. Experimental results show that neither feature encoding strategy reliably guided the agent toward feature-relevant code. The keyword-based approach produced feature representations that were too sparse, while the cluster-based approach generated overly dense connections that overwhelmed the agent with low-signal information. However, the use of graph-based tooling, particularly the ability to retrieve and inspect individual code components such as functions and classes, significantly improved the efficiency with which the agent explored the codebase. Additionally, token budgeting showed promise as a mechanism for limiting unnecessary code generation, though accurate budget estimation remains an open problem. Overall, this work highlights the limitations of explicit feature abstractions for agent understanding, while demonstrating the value of structural code representations and targeted retrieval. These findings suggest that agent systems benefit more from lightweight graph-based navigation and adaptive retrieval strategies than from feature encodings.
MIT Department
Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
