TLDR: This research paper introduces Structured Agentic Software Engineering (SASE), a framework for the new SE 3.0 era where intelligent agents handle complex software development tasks. It proposes a duality of ‘SE for Humans’ and ‘SE for Agents,’ reimagining traditional pillars like actors, processes, tools, and artifacts. Key components include specialized workbenches (Agent Command Environment for humans, Agent Execution Environment for agents) and structured, version-controlled artifacts (BriefingScript, LoopScript, MentorScript, Consultation Request Packs, Merge-Readiness Packs) to facilitate disciplined human-AI collaboration. The paper outlines engineering activities for this new paradigm and discusses its implications for software engineering education, emphasizing a shift from coding to agent orchestration and mentorship.
Software engineering is undergoing a profound transformation, moving beyond simple AI-assisted coding to an era where intelligent agents take on complex, goal-oriented objectives. This new phase, dubbed Agentic Software Engineering (SE 3.0), demands a fresh perspective on how humans and AI collaborate to build trustworthy software.
At the heart of this shift is a fundamental duality: Software Engineering for Humans (SE4H) and Software Engineering for Agents (SE4A). These two modalities work together, each with distinct needs and contributions. For humans, the role evolves into that of an ‘Agent Coach,’ focusing on high-level strategy, intent, and mentorship. For agents, it’s about creating a structured and predictable environment where they can operate effectively.
This new vision requires a complete rethinking of the traditional pillars of software engineering: the actors involved, the processes they follow, the tools they use, and the artifacts they create. In Agentic SE, the ‘actors’ expand from human developers to a hybrid team of human coaches and specialized AI agents. ‘Processes’ move from informal prompting to structured, repeatable engineering activities that govern human-agent collaboration. ‘Artifacts’ transform from transient prompts into durable, machine-readable documents that serve as contracts and institutional memory. Finally, ‘tools’ evolve from the traditional all-in-one human-centric Integrated Development Environment (IDE) to specialized workbenches tailored for both humans and agents.
Two key environments emerge from this duality: the Agent Command Environment (ACE) and the Agent Execution Environment (AEE). The ACE is the human coach’s command center, optimized for strategic tasks like orchestrating workflows, specifying intent, and reviewing agent-generated results. It provides full visibility into agent activities and costs. The AEE, on the other hand, is the agents’ digital workbench, designed for their unique strengths such as high-speed computation, massive parallelism, and tireless execution. This bi-directional partnership allows agents to proactively seek human expertise when facing complex decisions or ambiguities, and humans to oversee and guide agent teams.
The interaction between these environments is not a casual chat, but a structured dialogue managed through a series of explicit, version-controlled, and machine-readable artifacts. These include the BriefingScript, which outlines the mission plan; the LoopScript, defining the workflow playbook; and the MentorScript, a guide for best practices. Agents respond by generating their own formal artifacts, such as the Consultation Request Pack (CRP) to ask for human expertise, and the Merge-Readiness Pack (MRP) to present a final, evidence-backed deliverable. Humans then complete the loop with Version Controlled Resolutions (VCRs), which formally address these requests, ensuring a traceable and continuous conversation.
This structured approach gives rise to several key engineering activities. Briefing Engineering (BriefingEng) focuses on creating clear, actionable mission briefs for agents. Agentic Loop Engineering (ALE) governs how agents execute tasks, defining collaboration patterns and tool engagement. AI Teammate Mentorship Engineering (ATME) codifies team norms and best practices into ‘mentorship-as-code.’ Agentic Guidance Engineering (AGE) formalizes human intervention, ensuring efficient and targeted input. Finally, AI Teammate Lifecycle Engineering (ATLE) and AI Teammate Infrastructure Engineering (ATIE) build the foundation for agents to retain memory, learn over time, and operate effectively within their specialized environment.
The goal of this Structured Agentic Software Engineering (SASE) vision is not to offer a definitive solution, but to provide a conceptual framework and vocabulary to encourage dialogue within the software engineering community. It pushes us to think beyond human-centric tenets towards a disciplined, scalable, and trustworthy agentic future. This shift promises unprecedented productivity gains, potentially moving beyond the ’10x developer’ to ‘100x or even 1000x productivity,’ by redefining the role of developers from coders to orchestrators and mentors of AI teammates.
This new era also has profound implications for software engineering education. Instead of training students to be coders, we must prepare them to manage fleets of agents, emphasizing strategic skills like system-level thinking, architectural reasoning, rigorous specification, and the art of mentorship-as-code. This is a call for a holistic reimagining of what it means to educate the next generation of software engineers.
Also Read:
- Navigating the Decentralized World: The Role of AI Agents in Blockchain Security and Privacy
- HECATE: A Framework Bridging Multi-Agent Systems and Distributed Systems Engineering
For a deeper dive into the foundational pillars and research roadmap of this exciting new field, you can read the full research paper: Agentic Software Engineering: Foundational Pillars and a Research Roadmap.
