Agentic AI applications in software development : a systematic mapping study
Khan, Feroz (2026)
Diplomityö
2026
School of Engineering Science, Tietotekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2026042433968
https://urn.fi/URN:NBN:fi-fe2026042433968
Tiivistelmä
The rapid development of generative AI technologies has impacted software engineering, marking a shift from responsive to autonomous and goal-oriented systems which we refer to as Agentic AI. This paper provides a systematic mapping study (SMS) of Agentic AI through its various aspects according to the SDLC framework. Following the methodology established by Peterson et al. (2008), 20 primary sources on the topic dated between 2025 and 2026 were reviewed, and the findings have been classified into five key technological dimensions.
In SDLC, applications of Agentic AI remain concentrated in later stages, particularly coding, testing, DevOps, and maintenance, with early phases such as requirements and design still underexplored. From an architectural perspective, the field is shifting toward distributed and collaborative approaches, with multi-agent systems and orchestrator–worker models emerging as dominant patterns. The technology stack reflects a layered ecosystem, where large language models combined with reinforcement learning form the foundation. Coordination between agents is facilitated through shared memory and standardised protocols.
In addition to the technical aspects, the study also identifies socio-technical challenges associated with the adoption of agentic software, specifically issues related to reliability and trust. Evaluation practices remain fragmented and task-specific, limiting comparability across studies.
In summary, the study shows that agentic AI is developing into a cohesive approach whereby software development takes place through systemic transformation. This study provides a taxonomy for the study of agentic AI and opens up possibilities for future research, including the establishment of standard evaluation techniques, improved approaches to human and AI cooperation, and creating agentic AI systems.
In SDLC, applications of Agentic AI remain concentrated in later stages, particularly coding, testing, DevOps, and maintenance, with early phases such as requirements and design still underexplored. From an architectural perspective, the field is shifting toward distributed and collaborative approaches, with multi-agent systems and orchestrator–worker models emerging as dominant patterns. The technology stack reflects a layered ecosystem, where large language models combined with reinforcement learning form the foundation. Coordination between agents is facilitated through shared memory and standardised protocols.
In addition to the technical aspects, the study also identifies socio-technical challenges associated with the adoption of agentic software, specifically issues related to reliability and trust. Evaluation practices remain fragmented and task-specific, limiting comparability across studies.
In summary, the study shows that agentic AI is developing into a cohesive approach whereby software development takes place through systemic transformation. This study provides a taxonomy for the study of agentic AI and opens up possibilities for future research, including the establishment of standard evaluation techniques, improved approaches to human and AI cooperation, and creating agentic AI systems.