Modern software systems operate in environments where workloads, dependencies, and execution conditions cannot be fully anticipated at design time. Self-adaptive systems address such uncertainty by adapting their behavior or structure at runtime. However, their adaptation mechanisms still largely rely on knowledge, rules, and capabilities defined before deployment. Recent advances in generative and agentic AI provide new capabilities for reasoning, tool use, and collaboration, opening new possibilities for software adaptation while raising challenges related to the organization, configuration, and control of agentic systems. This thesis investigates how agentic AI can support the runtime adaptation of complex software systems and how the agentic systems supporting this adaptation can themselves be organized and engineered. It addresses this problem through four complementary contributions. First, a systematic mapping characterizes the use of generative AI for adapting microservice architectures. From 2,740 candidate publications, 22 studies were selected. The results show a predominant focus on reactive adaptation, while proactive adaptation and industrial validation remain less explored. Second, AWARE (Assess, Weigh, Act, Reflect, and Enrich) introduces an agent-oriented organization of the managing system in which adaptation responsibilities are distributed among specialized agents. Third, the Assess capability of AWARE is instantiated for root cause analysis in microservice architectures through a dynamic agent-generation approach that adapts the organization of diagnostic agents to each incident. Evaluations with seven large language models on the Nezha and OpenRCA datasets show improvements of up to 12% and 22%, respectively, compared with the reported baselines. These results indicate that the organization of the managing system itself can become an adaptation dimension. Finally, GEAR introduces a feature-model-based approach for specifying and configuring AI-based multi-agent systems independently of the frameworks used to implement them. To extend its transformation capabilities, we examined how each of the 32 features of the GEARWorkflow model could be translated into implementations across ten agentic frameworks. This resulted in 320 comparisons, with clear mappings identified in 271 cases (84.69%). These mappings provide a basis for developing and extending the framework-specific connectors of GEAR. Together, these contributions progress from understanding how generative capabilities support software adaptation, to organizing them through agents, dynamically adapting this organization, and engineering its configuration variability. They show that agentic software adaptation concerns not only the managed system, but also the organization of the managing system itself, whose increasing flexibility requires explicit representations, constraints, and validation mechanisms to remain controllable and reproducible.
- M. Clément QUINTON Professeur des universités Université de Lille Directeur de thèse, - M. Olivier BARAIS Professeur des universités Université de Rennes Examinateur, - Mme Amel BENNACEUR Professeure The Open University Rapporteure, - M. Alain TCHANA Professeur des universités Université Grenoble INP – UGA Rapporteur, - M. Romain ROBBES Directeur de recherche CNRS Examinateur, - M. Paul TEMPLE Maître de conférences Université de Rennes Co-encadrant de thèse, - Mme Anne ETIEN Professeure des universités Université de Lille Examinatrice.
Thesis of the team Spirals defended on 03/12/2026