AgileCoder is a multi-agent software development framework that adapts Agile Methodology to LLM agents; its repository is given as https://github.com/FSoft-AI4Code/AgileCoder. Given user requirements, agents in Agile roles plan, write, review and test the software incrementally over sprints, with the framework deciding at the end of each sprint whether to deliver or plan another. It is introduced in AgileCoder: Dynamic Collaborative Agents for Software Development based on Agile Methodology, whose authors present it as an alternative to multi-agent systems such as MetaGPT and ChatDev, which they describe as following the waterfall model.
Capabilities
The Product Manager turns user requirements into a product backlog of development tasks and acceptance criteria, and the Scrum Master reviews it for feasibility. Each sprint then runs four phases. In planning, tasks are selected into a sprint backlog. In development, the Developer writes code annotated with docstrings, and the Senior Developer reviews it statically in three steps: basic implementation checks such as empty methods and missing imports, compliance with the sprint backlog, and satisfaction of the acceptance criteria with no bugs. In testing, the Tester writes test suites for the files changed in the sprint and their ancestor files, and the files are executed according to a testing plan until bugs or failed tests are reported back to the Developer. In review, the Product Manager compares an accumulated report with the backlog and acceptance criteria to decide whether to conclude, in which case the Scrum Master writes documentation on how to run the software and install its libraries.
The Dynamic Code Graph Generator keeps a Code Dependency Graph whose nodes are code files and whose edges mainly capture import relations, updating it as the code changes. The graph determines which files need testing, supplies a logical testing order by reversing a topological order, and, when an execution error occurs, lets agents trace back from the traceback and retrieve only the relevant cross-file context instead of the whole codebase. An Execution Environment runs the code and returns tracebacks to the agents.
Agents talk in two-party conversations between an instructor and an assistant in unconstrained natural language, continuing until they reach consensus or an exchange limit; a message stream serves as the conversation’s working memory, and a global message pool stores every conversation’s output and the status of each task, from which each conversation reads only the segments relevant to it.
Adoption & Ecosystem
The paper runs AgileCoder with GPT-3.5 Turbo, Claude 3 Haiku and GPT-4 as backbone models and evaluates it on HumanEval, MBPP and ProjectDev, a set of software development tasks the authors compiled. The authors suggest extending it with further Agile practices such as pair programming, CI/CD, Kanban or Lean, and applying the approach to domains beyond software development.