“Agentic Software Engineering: Foundational Pillars and a Research Roadmap” is a 2025 vision paper by researchers at Queen’s University, Concordia University, the Nara Institute of Science and Technology, and Huawei Canada. It argues that Agentic Software Engineering (SE3.0) needs disciplined structure rather than just more capable agents, and introduces a duality between SE for Humans and SE for Agents that reimagines software engineering’s actors, processes, tools, and artifacts around that split. Its central proposal is Structured Agentic Software Engineering (SASE) (SASE), a conceptual scaffold of dedicated environments, version-controlled artifacts, and named engineering activities for the two modalities. The paper also proposes a six-level hierarchical framework distinguishing agency from autonomy, grounds its argument in a worked example of a developer resolving seven pull requests through a team of agents, and closes with a research roadmap and a discussion of implications for software engineering education.
Key Points
- Proposes Structured Agentic Software Engineering (SASE) (SASE), reimagining software engineering’s four pillars — actors, processes, tools, and artifacts — around a duality of SE for Humans (SE4H) and SE for Agents (SE4A).
- Introduces SE Autonomy Levels (SE1.0–SE5.0), a six-level hierarchy (SE1.0–SE5.0) distinguishing agency (executing a plan) from autonomy (formulating goals), modeled on SAE self-driving automation levels.
- Defines two purpose-built workbenches — the Agent Command Environment (ACE) (ACE) for human coaches and the Agent Execution Environment (AEE) (AEE) for agents — connected by version-controlled artifacts such as BriefingScript, LoopScript, and MentorScript.
- Names five structured engineering activities — Briefing Engineering (BriefingEng), Agentic Loop Engineering (ALE), AI Teammate Mentorship Engineering (ATME), Agentic Guidance Engineering (AGE), and a joint AI Teammate Lifecycle Engineering (ATLE)/AI Teammate Infrastructure Engineering (ATIE) discipline — that operationalize SASE.
- Reports that recent audits of SWE-Bench-style benchmark results show agent-generated code frequently falls short of being merge-ready despite passing tests, citing findings such as GPT-4 patch true-solve rates dropping from 12.47% to 3.97% after manual review, and traces a progression from SWE-Bench to SWE-Bench Verified (introduced to address task ambiguity and underspecification) to SWE-Bench Pro (recommended as contamination concerns grew).
- Cites an early study finding that 83.8% of Claude Code-assisted pull requests on GitHub were eventually merged, and cites the AIDev dataset of 932,791 agent-authored pull requests as evidence that agentic coding is already widespread.
Notes
The paper positions itself as a “conceptual scaffold,” not a definitive solution, explicitly inviting the community to challenge, refine, and extend its proposed activities. It compares SASE to related efforts — the Plan-Do-Assess-Review (PDAR) loop with Product Requirement Prompt (PRP)s, agent-skill plugin libraries, meta-prompt files like CLAUDE.md/AGENT.md (see AGENTS.md), and the BMAD (Breakthrough Method for Agile AI-Driven Development) multi-agent framework — arguing SASE differentiates itself through “mentorship-as-code,” dual workbenches, merge-readiness as the target artifact, and consultation as a first-class artifact. It builds on the same authors’ earlier “Towards AI-Native Software Engineering (SE3.0)” proposal, cited in the paper’s own references but not itself analyzed here.