Agentic AI for Software Teams: A Research Collaboration with Atlassian

This is a collaborative research project with Atlassian. Check out some recent work:

- https://www.atlassian.com/software/rovo-dev

- https://www.atlassian.com/blog/atlassian-engineering/hula-blog-autodev-paper-human-in-the-loop-software-development-agents

- https://arxiv.org/abs/2411.12924

🎯 Research Vision

The next generation of software engineering tools will move beyond autocomplete and static code generation toward autonomous, agentic systems — AI developers capable of planning, reasoning, and improving software iteratively. This project explores the development of agentic AI systems that act as intelligent collaborators: understanding project goals, decomposing problems, writing and testing code, and learning from feedback.

🔍 Research Objectives

1. Design an Agentic SWE Framework

   • Model an AI system that combines reasoning, planning, and self-correction for software engineering tasks.

   • Integrate components such as task decomposition, code synthesis, test generation, and evaluation.

2. Evaluate Agentic Behavior in Code Generation

   • Assess how agentic systems (e.g., those using LLMs with reflection or reinforcement loops) compare to static code generators.

   • Metrics: code quality, correctness, maintainability, and speed of convergence.

3. Develop Benchmark Tasks for AI Developers

   • Curate open-source software tasks (bug fixing, refactoring, unit test creation) for empirical evaluation.

   • Measure human-AI collaboration efficiency.

4. Investigate Trust and Explainability

   • Explore how developers interact with autonomous AI coders.

   • Build explainability modules to visualize agent reasoning and decisions.

⚙️ Methodology

• LLM Foundation: Use open-source models (e.g., Code Llama, DeepSeek-Coder, or GPT-4o-mini).

• Agentic Loop: Implement reflection–action cycles using frameworks such as LangChain, AutoGen, or CrewAI.

• Evaluation:

  • Static analysis (code correctness, test coverage).

  • Developer study (usability and trust).

  • Learning curves over multiple development iterations.

🧪 Expected Outcomes

• Prototype of an agentic code generation framework capable of self-directed code improvement.

• Empirical evidence of productivity gains and trust challenges.

• Publications in venues like ASE, ICSE, or ESEC/FSE (for advanced master’s work).

🔮 Future Impact

This research contributes to the emerging field of Agentic Software Engineering (Agentic SWE) — redefining what it means to “develop software.” Students will gain exposure to autonomous AI systems, software intelligence architectures, and human-in-the-loop development — core technologies for the next decade of AI-powered engineering.