Human in the Loop Software Development Agents

This post introduces HULA (Human-in-the-loop LLM-based agents framework) which automates repetitive development tasks by reading Jira work items, generating coding plans, writing code, and raising pull requests. It details the HULA process (Set context, Generate coding plan, Generate code, Raise pull request) and presents performance metrics including SWE-bench unit test pass rates (31%), GPT-based similarity to human code (45%), engineer usage statistics (79% plan generation approval, 87% code generation from approved plans, 25% to PR stage, 59% merged PRs), and survey feedback on plan alignment (36%), code understandability (61%), and code solution effectiveness (33%).