AI Agents for Kubernetes Incident Remediation: Evaluating the Reflexion Framework

---

Thesis Type: Postgraduate

Institution Of The Thesis: TED University, Graduate School, Computer Engineering, Turkey

Approval Date: 2026

Thesis Language: English

Student: MUSTAFA MERT SÜERKAN

Principal Supervisor (For Co-Supervisor Theses): Emin Kuğu

Co-Supervisor: Hakan Emekci

Open Archive Collection: AVESIS Open Access Collection

Abstract:

Kubernetes has become a standard technology used in many infrastructures worldwide for container orchestration. However, troubleshooting errors in a Kubernetes environment requires expert knowledge. Existing solutions within Kubernetes, such as Horizontal Pod Autoscaler and Vertical Pod Autoscaler, offer limited autonomous correction capabilities. This thesis analyzes the effectiveness of the Reflexion Framework, which uses a verbal reinforcement learning approach to autonomously troubleshoot problems in a Kubernetes environment. During the experiment, 1,500 structured tests were completed with five Claude models, ten troubleshooting scenarios, and three experimental configurations (baseline, two-attempt, and full Reflexion). The results show that allowing the agents a second attempt increases the success rate from 38.6% to 68.4%. The addition of the Reflexion component results in a large increase of 12 percentage points for troubleshooting situations involving several failures that need to be treated together but shows little improvement for situations involving a single failure. The capacity of the models was also an important determinant of the Reflexion results, with the strongest results for the medium-capacity models; the low-capacity models were unable to make effective use of the Reflexion content. Strong models show ceiling effects that achieve high success rates without requiring reflection because high-capability models already succeeded on the second attempt without explicit reflection. Considering speed, success, and cost, medium capacity models were observed as the optimal solution. These results show that Reflexion is not always beneficial, but it provides significant value when used in complex problems with appropriate model selection.