AgileCoder: Dynamic Collaborative Agents for Software Development based on Agile Methodology (2406.11912v2)

Abstract: Software agents have emerged as promising tools for addressing complex software engineering tasks. Existing works, on the other hand, frequently oversimplify software development workflows, despite the fact that such workflows are typically more complex in the real world. Thus, we propose AgileCoder, a multi agent system that integrates Agile Methodology (AM) into the framework. This system assigns specific AM roles - such as Product Manager, Developer, and Tester to different agents, who then collaboratively develop software based on user inputs. AgileCoder enhances development efficiency by organizing work into sprints, focusing on incrementally developing software through sprints. Additionally, we introduce Dynamic Code Graph Generator, a module that creates a Code Dependency Graph dynamically as updates are made to the codebase. This allows agents to better comprehend the codebase, leading to more precise code generation and modifications throughout the software development process. AgileCoder surpasses existing benchmarks, like ChatDev and MetaGPT, establishing a new standard and showcasing the capabilities of multi agent systems in advanced software engineering environments.

Dynamic Collaborative Agents for Software Development Using Agile Methodology

The paper introduces AgileCoder, a multi-agent system designed to enhance software development by integrating Agile Methodology (AM) roles within its framework. This approach assigns distinct AM roles such as Product Manager, Developer, and Tester to different agents, enabling a collaborative software development process based on user inputs. AgileCoder organizes work into iterative sprints, improving development efficiency through a focus on incremental progress. A significant innovation is the Dynamic Code Graph Generator, which dynamically creates a Code Dependency Graph as the codebase updates, enhancing agents' understanding and enabling precise code modifications.

Background and Motivation

The advent of autonomous software agents leveraging LLMs offers considerable promise for improving software development workflows. Previous systems, such as MetaGPT, which encodes SOPs into tasks, and ChatDev, a virtual chat-powered technology company, often simplify workflows using the waterfall model. However, real-world development is predominantly Agile (about 70% of professional teams use AM), which emphasizes adaptability and iterative enhancement. Existing models overly depend on LLMs for decision-making in complex code generation tasks. AgileCoder's design directly addresses these oversimplifications by embedding Agile's iterative nature within a multi-agent framework.

System Design and Implementation

AgileCoder employs a comprehensive set of agent roles that simulate a professional Agile development environment. These roles include:

• Product Manager (PM): Curates product backlog tasks based on user requirements.
• Scrum Master (SM): Ensures sprint backlog feasibility and manages development processes.
• Developer (Dev): Focuses on code generation and refactoring.
• Senior Developer (SD): Conducts reviews and quality control.
• Tester: Creates and executes test cases aligned with development requirements.

The development workflow comprises several phases: Planning, Development, Testing, and Review. Each of these phases is strategically implemented to enhance software development, with dynamic agent interactions mirroring professional team operations.

Key Innovations

The introduction of the Dynamic Code Graph Generator (DCGG) is pivotal. DCGG provides a Code Dependency Graph (CDG) that dynamically updates to reflect evolving codebase relationships. This structure supports agents in retrieving context-specific data, ensuring that updates and bug fixes are informed by current code dependencies, thus avoiding the typical inefficiencies of static code analysis.

Experimental Evaluation

Experiments conducted on benchmarks such as HumanEval and MBPP, along with a new ProjectDev dataset, demonstrate AgileCoder's efficacy. AgileCoder surpasses existing systems like ChatDev and MetaGPT, achieving state-of-the-art performance in code generation tasks. Notably, it achieves a pass@1 of 70.53% on HumanEval and 80.92% on MBPP using GPT-3.5 Turbo, marking substantial improvements over previous benchmarks.

Implications and Future Work

AgileCoder's framework offers both practical enhancements and theoretical insights into software engineering automation. The integration of Agile Methodology principles into multi-agent systems could revolutionize collaborative software development, emulating human development processes more accurately than traditional waterfall-inspired models.

Future developments might include extending AgileCoder's capabilities beyond software development to areas such as product design and project management. Moreover, refining LLMs' robustness in code generation and ensuring scalability for large-scale projects remain essential areas for research. Further incorporation of Agile practices like pair programming and CI/CD would enhance its application breadth.

Overall, AgileCoder marks a significant contribution to the field of software engineering automation, leveraging multi-agent frameworks to emulate complex human development workflows effectively.