บทวิเคราะห์ paper 'Agentic Software', วิชาที่เกิดใหม่เมื่อ agent เข้ามาแทนที่โค้ด

The End of Software Engineering: How AI Agents Are Restructuring the Software Paradigm, a recent paper by Zhenfeng Cao from Lingxi Intelligent Investment, explores how AI agents are transforming the field of software engineering. Traditionally, software engineering has been focused on breaking down complex problems into small, manageable pieces of code, which are then maintained over time.

However, with the advent of AI agents, particularly those based on large language models (LLMs), the role of code is shifting towards being a temporary tool for reasoning rather than the primary product itself.

The paper outlines three phases in the evolution of software development:

1. The early phase, marked by the creation of software engineering in response to the software crisis in the 1960s, where the focus was on breaking down problems into small, logical pieces of code and maintaining them over time. This approach proved successful for many decades, thanks to Fred Brooks' observation that the complexity of software increases exponentially due to the interconnectedness of features.

2. The second phase, characterized by the shift from on-premise software to cloud-based (SaaS) and agent-as-a-service (AaaS) models. In AaaS, users simply specify their desired outcomes, and the agent handles the rest, transforming the role of code from a system to a temporary means of reasoning.

3. The third phase, which is still developing, involves the emergence of Agentic Engineering. This new discipline revolves around designing autonomous systems of AI agents that work together, much like a team, to design, develop, and refine software systems without constant human intervention.

The authors suggest that while we may currently be at the third stage, true autonomy in software engineering is still a long way off. The paper also presents a roadmap for the future development of software agents, highlighting the potential for self-evolving ecosystems where AI agents can evolve their own design principles and spawn new generations of agents without human oversight.

However, the authors acknowledge that while the shift towards agent-based software paradigms is significant, human oversight will still be necessary, particularly in high-stakes or complex projects. They note that while AI agents can significantly improve efficiency and reduce the burden of maintaining legacy code, the full potential of truly autonomous software systems is still a theoretical goal.