- The paper's main contribution is the systematic identification and prioritization of research directions across six thematic areas related to agentic AI in software engineering.
- The methodology combined expert elicitation, collaborative clustering, and LLM-augmented thematic coding to rigorously analyze both technical and sociotechnical aspects.
- Key implications include embedded governance models, agent-augmented architectural practices, and new metrics for quality and sustainability in software development.
A Research Agenda at the Intersection of Agentic AI and Software Engineering
Introduction
The progression of agentic AI constitutes a profound transformation within software engineering, compelling the discipline to contend with both the integration of autonomous agents into development workflows and the necessity of robust engineering principles for the construction of agentic systems. The paper "A Research Agenda on Agents and Software Engineering: Outcomes from the Rio A2SE Seminar" (2605.11720) summarizes the collective scrutiny and discussion of 18 distinguished participants from academia and industry, resulting in a systematically prioritized agenda addressing both immediate and future research imperatives. The effort is distinctive in framing the multifaceted challenges presented by agentic AI in six carefully delineated thematic areas.
Methodology and Seminar Design
The research agenda derives from a multi-phase seminar structure involving expert elicitation, collaborative topic clustering, parallel group discussions, and rigorous qualitative analysis augmented by LLM tools for thematic coding. Participants—spanning architecture, quality, sustainability, and knowledge transfer across industry and academia—deliberated on both the technical and sociotechnical properties of agentic integration within software engineering. The six emergent thematic compass points—Governance, Software Engineering for Agents (SE4A), Agents for Software Architecture (A4SA), Quality and Evaluation, Sustainability, and Code—emerged through consensus and iterative validation.
Thematic Areas and Prioritized Research Directions
Governance
Short-term directions target operationalizing responsibility, compliance, and transparency via agent governance models, policy integration, and DevOps adaptation (AgentOps) for safe deployment. Longer-term objectives advance towards the institutionalization of governance, including certification mechanisms, multi-layer frameworks, and the formalization of governance debt. The research underscores the importance of embedding governance into system design rather than externalizing it, marking a notable shift in accountability paradigms for agentic systems.
Software Engineering for Agents
The agenda advocates immediately for hybrid and orchestrated architectures, explicit design granularity, and accelerated framework development for agent-based systems. Long-term research must address technology-agnostic methodologies, empirical knowledge consolidation, formal verification, and the articulation of design patterns specific to non-deterministic agent behaviors. The explicit call for community-driven repositories and technology-agnostic guidance reflects the pressing need for standardization amid rapidly diversifying agentic technologies.
Agents for Software Architecture
Near-term priorities include specification-driven practices and agent augmentation of human architectural decision-making, supported by systematic benchmarking of agents on architectural tasks. Long-term implications propose a redefinition of core architectural concepts—most prominently, maintainability—given emergent trends such as code being superseded by specifications and increased automation of architectural knowledge management and analysis. The evolution of architectural education is also highlighted as a critical vector for enabling practitioners in AI-driven environments.
Quality and Evaluation
Quality assurance is foregrounded in short-term research on robustness, bias mitigation, and securing agentic systems. The long-term vision expands towards new agent-specific quality metrics, ecosystem robustness (including agent-generated agents), and the development of both human-readable and machine-readable code artifacts. Verification of AI-generated outputs and the anticipated transformation of coding practices and maintainability models are emphasized as requiring significant foundational work.
Sustainability
Resource awareness, sustainability metrics, and explicit ROI models guide short-term agendas, with an acute focus on energy consumption, long-term maintainability, and heightened user awareness of sustainability trade-offs. In the long term, the transition to hardware-agnostic benchmarking and large-scale modernization toward sustainable, agent-oriented architectures is identified as central, reflecting the dual environmental and economic imperatives facing the adoption of large-scale agentic software systems.
Code
Immediate research foci include the quality of generated code, technical debt, and the reverse engineering capacity of agents. Architectural modernization via agents is also central. Over longer horizons, the agenda emphasizes robust system–agent communication protocols, runtime code generation, evolved maintainability models for continuously generated code, and frameworks for cost-based and context-aware decisions between agentic and traditional development paradigms. The prospect of runtime auto-generation and self-modifying code is a salient, if complex, research trajectory.
Implications and Prospective Developments
The agenda demonstrates clear convergence on several cross-cutting challenges:
- Canonical frameworks and reusable patterns: There is a compelling requirement for standardization, empirical evaluation, and community-driven repositories to coordinate the rapidly growing but fragmented landscape of agentic engineering practices.
- Embedded governance and policy: The integration of accountability, compliance, and explainability as first-class concerns marks a shift toward more trustworthy and governable agentic systems.
- Sociotechnical transformation: The rise of agentic AI necessitates a recalibration of human-AI collaboration, with new roles and engineering competencies emerging and evolving.
- Quality, sustainability, and maintainability: There is a recognized need for new metrics, modeling approaches, and empirical methodologies attuned to the characteristics of autonomous, non-deterministic systems.
These directions foreshadow heightened integration of empirical research (including agents for science), increased interaction between academia and industry, and a holistic approach that recognizes the inseparability of technical and organizational aspects in agentic AI adoption.
Conclusion
The research agenda outlined in "A Research Agenda on Agents and Software Engineering: Outcomes from the Rio A2SE Seminar" provides a rigorously developed, multi-temporal framework anchoring future investigation at the intersection of agentic AI and software engineering. By foregrounding non-determinism, specification-driven practices, and the expanding scope of code generation and system modernization, the agenda confronts the theoretical and practical recalibration demanded by agent integration. The explicit focus on empirical validation, community standards, and continuous knowledge consolidation will significantly shape both the immediate evolution and long-term maturation of agentic software engineering.