The rise of large language model (LLM) agents is reshaping how software is built, promising near-complete automation from natural language requirements. Yet, most current approaches rely on linear, waterfall-style pipelines that struggle to capture the iterative and interconnected nature of real-world projects—especially complex or large-scale software.
To address these limitations, researchers Junwei Liu, Chen Xu, Chong Wang, Tong Bai, Weitong Chen, Kaseng Wong, Yiling Lou, and Xin Peng introduced EvoDev, a novel framework inspired by feature-driven development. Rather than processing requirements sequentially, EvoDev breaks down user needs into discrete, value-focused features. These features are organized into a Feature Map, a directed acyclic graph that explicitly represents dependencies between features.
Each feature node in EvoDev stores multi-level information, including business logic, design, and code. This structured context is propagated along dependencies, ensuring that subsequent development iterations have full visibility into the work completed so far. The approach mirrors real-world software development more closely than linear pipelines, allowing LLM agents to iterate, adapt, and build complex software in a coherent, dependency-aware manner.
Evaluating EvoDev’s Impact
The EvoDev team tested the framework on challenging Android development tasks. Results were striking: EvoDev outperformed the best-performing baseline, Claude Code, by 56.8%. Single-agent performance also improved across different LLM bases, ranging from 16% to 76.6%, demonstrating the power of dependency modeling, context propagation, and iterative workflows.
These findings underscore the critical role of workflow-aware agent design and dependency-aware context propagation in handling sophisticated software projects. EvoDev not only boosts efficiency but also provides actionable insights for future LLM training, helping AI agents better understand and support iterative software development.
Implications for the Future
EvoDev signals a shift in AI-assisted development toward iterative, feature-driven frameworks. By enabling LLMs to understand project dependencies and propagate knowledge across iterations, developers can tackle larger, more complex projects with improved accuracy and reduced rework.
The framework’s insights extend beyond Android development, offering guidance for designing future AI-driven software tools. As LLMs continue to evolve, frameworks like EvoDev could become the standard for intelligent, iterative software creation—merging human oversight with AI’s efficiency for end-to-end development.
