TLDR: A new research paper by Rashina Hoda advocates for expanding the vision of Agentic Software Engineering (SE) beyond just coding to a ‘whole of process’ approach. It proposes a framework encompassing ethical alignment, requirements, design, development, and operations, guided by CRAFT values (Comprehensive, Responsible, Adaptive, Foundational, Translational). The paper also emphasizes the critical need for a well-defined vocabulary to foster clear communication and collaboration in this emerging field, urging the SE community to adopt a systematic and principled approach to integrating AI agents.
The world of Software Engineering (SE) is on the brink of a significant transformation, driven by the rapid advancements in Agentic AI. While much of the current excitement around Agentic AI in software development focuses on automating coding tasks, a new research paper argues for a much broader perspective. Authored by Rashina Hoda from Monash University, the paper titled “Toward Agentic Software Engineering Beyond Code: Framing Vision, Values, and Vocabulary” proposes an expanded vision for Agentic SE, moving beyond just code to encompass the entire software development process.
Historically, Software Engineering has always been more than just programming. From the early days of the Waterfall model to the rise of Agile methodologies, SE has involved a structured approach to requirements, design, testing, and operations, supported by various tools, techniques, and roles. The paper highlights that current Agentic AI systems, such as SWE-agent and Devin, are primarily focused on code-related activities. However, early empirical evidence suggests that for Agentic AI to truly succeed in practice, it must address a wider range of socio-technical concerns, including teamwork, coordination, accountability, and human-AI collaboration.
The paper introduces a preliminary ‘whole of process’ vision for Agentic SE. This vision extends across several high-level core areas, where both human and AI agent actors collaborate iteratively. These areas include:
Agentic SE Ethical Alignment
This new area focuses on ensuring that AI and software ethics are integrated into the development process. With AI models gaining more autonomy and potential for widespread impact, ethical considerations like accountability, reliability, safety, transparency, and privacy become paramount.
Agentic Requirements Engineering
This involves the collaboration of humans and agents in defining software requirements, including customer needs, user expectations, quality standards, and ethical considerations.
Agentic Design
Here, agents assist in various design activities, from architectural decisions to user interface (UI) and user experience (UX) design.
Agentic Development
This is the area where most current Agentic AI efforts are concentrated, covering code design, production, testing, reviewing, and repairing. The paper acknowledges the progress here but emphasizes the need to integrate it with other phases.
Also Read:
- The Rise of Self-Governing AI: From External Control to Intrinsic Intelligence
- The Looming Divide: How Autonomous AI Agents Could Create New Societal Disparities
Agentic Operations
This final area encompasses activities related to deployment, maintenance, and ongoing operations, including practices like DevOps and DevSecOps.
To guide this paradigm shift, the paper proposes a preliminary set of CRAFT values and principles for Agentic SE:
- Comprehensive: Agentic SE should adopt a ‘whole of process’ approach, considering human and socio-technical aspects alongside technological innovation.
- Responsible: It must incorporate ethics-by-design and ensure sustainability, questioning the ‘why’ behind automation rather than just the ‘why not’.
- Adaptive: Agentic SE needs to remain relevant by adapting to new AI models and paradigms, working closely with the AI community and industry, and predicting socio-technical impacts.
- Foundational: The field requires developing foundational knowledge through empirical studies and creating foundational solutions that extend beyond just programming.
- Translational: Research insights must be translated into practice through awareness campaigns, education, and actionable guidelines and tools.
Finally, the paper stresses the importance of a well-defined vocabulary for Agentic SE. It highlights current discrepancies in terminology and offers considerations for designing and using terms, focusing on relevance, coverage, community acceptance, consistency, and philosophical alignment. For instance, it questions whether an agent performing only programming tasks should be termed a ‘software engineer’ or if ‘autonomous coding agent’ is more accurate.
This research paper serves as a call to action for the Software Engineering community to collaborate and lay strong, deliberate, and desirable foundations for Agentic SE, ensuring its evolution is systematic and principled. You can read the full paper here.
