Auto-Devops GPT: An Agentic AI Framework for Self-Healing CI/CD Pipelines Using LLM-Based Root Cause Analysis and Reinforcement Learning

Abstract

Continuous Integration and Continuous Deployment (CI/CD) pipelines are essential for modern software delivery, but frequent failures in build, test, dependency, configuration and deployment stages reduce reliability and increase recovery time. Traditional CI/CD tools execute predefined workflows but generally require manual log inspection and human intervention after failures. This paper proposes AutoDevOps GPT, an agentic artificial intelligence framework for self-healing CI/CD pipelines. The framework integrates Jenkins webhook events, log-based failure classification, Large Language Model-assisted root cause analysis, multi-agent recovery planning, safe remediation execution, incident storage and reinforcement learning-based recovery optimisation. The proposed system uses an Analyzer Agent to classify failures, a Planner Agent to select recovery actions and an Executor Agent to apply predefined safe remediation. A lightweight Q-learning-inspired policy updates recovery action preferences using success, time and cost feedback. The system is implemented using Python Flask, SQLite, SQLAlchemy, Jenkins, GitHub, Docker, Kubernetes support and a dashboard for incident intelligence. Experimental evaluation shows improvements in detection time, recovery time, recovery accuracy and automation efficiency compared with traditional CI/CD, rule-based recovery and basic AIOps baselines.

Introduction

This paper presents AutoDevOps GPT, an agentic AI-based self-healing framework for Continuous Integration and Continuous Deployment (CI/CD) pipelines. Although CI/CD automates software development tasks such as building, testing, packaging, and deployment, pipeline failures caused by dependency issues, build errors, configuration mistakes, container failures, or Kubernetes deployment problems often require manual intervention. These delays increase Mean Time to Recovery (MTTR) and reduce development productivity.

To address these challenges, AutoDevOps GPT integrates rule-based failure detection, LLM-assisted root cause analysis, multi-agent recovery planning, safe automated remediation, and reinforcement learning-inspired optimization into a closed-loop recovery system. Unlike traditional CI/CD and many AIOps tools, the framework not only detects failures but also diagnoses, plans, executes, and continuously improves recovery strategies.

Key Contributions

• An agentic AI architecture for automated CI/CD failure diagnosis and recovery.
• A hybrid failure analysis combining rule-based log matching with LLM reasoning.
• A safe recovery mechanism that executes only predefined and approved remediation actions.
• A Q-learning-inspired policy that improves recovery decisions using feedback.
• A Flask-based implementation integrated with Jenkins, GitHub, SQLite, Docker, and Kubernetes.
• Experimental comparison with traditional CI/CD, rule-based recovery, and AIOps approaches.

Framework Architecture

The proposed system consists of five layers:

1. Source Control (GitHub)
2. CI/CD Orchestration (Jenkins)
3. Monitoring Layer
4. Agentic AI Layer (Analyzer Agent, Planner Agent, Executor Agent, and LLM)
5. Feedback and Learning Layer (SQLite database and reinforcement learning policy)

Each agent has a specific role:

• Analyzer Agent: Detects failure types and identifies root causes.
• Planner Agent: Chooses the most suitable recovery action while enforcing safety constraints.
• Executor Agent: Performs approved recovery actions such as retrying pipelines, fixing dependencies, rolling back deployments, or creating GitHub issues.

Methodology

The framework follows a closed-loop workflow:

1. Detect pipeline failures using Jenkins webhooks.
2. Preprocess logs and classify failures through pattern matching.
3. Use an LLM to analyze unknown or complex errors.
4. Generate structured root cause reports with confidence scores.
5. Select safe recovery actions using learned policy values.
6. Execute only whitelisted remediation actions.
7. Compute rewards based on recovery success, time, and operational cost.
8. Update recovery policies using a Q-learning-inspired algorithm to improve future decisions.

Implementation

The prototype was developed using:

• Python Flask backend
• SQLite with SQLAlchemy
• Jenkins for CI/CD events
• GitHub for issue creation
• Docker and Kubernetes support
• HTML/CSS/JavaScript dashboard
• Simulation mode for evaluation without production infrastructure

The dashboard displays incident statistics, resolved and pending failures, recovery rewards, and recent pipeline failures.

Experimental Results

The framework was evaluated on common CI/CD failure scenarios, including dependency, build, test, timeout, deployment, configuration, and unknown failures. Compared with traditional CI/CD, rule-based recovery, and AIOps systems, AutoDevOps GPT achieved significant improvements:

• Failure Detection Time: Reduced from 120 seconds to 35 seconds
• Mean Time to Recovery (MTTR): Reduced from 25 minutes to 8 minutes
• Pipeline Success Rate: Increased from 78% to 94%
• Resource Utilization: Improved from 65% to 82%
• Recovery Accuracy: Increased from 68% to 93%
• Automation Efficiency: Improved from 35% to 88%

These improvements result from webhook-driven monitoring, intelligent agent collaboration, LLM-assisted reasoning, safe automated recovery, and continuous learning from previous incidents.

Conclusion

This paper presented AutoDevOps GPT, an agentic AI frame-work for self-healing CI/CD pipelines. The framework inte-grates Jenkins failure detection, log classification, LLM-based root cause analysis, multi-agent recovery planning, safe reme-diation execution, incident storage and reinforcement learning-inspired policy update. The system is implemented using Flask, SQLite, Jenkins, GitHub, Docker and Kubernetes sup-port. Experimental results show that AutoDevOps GPT improves all evaluated metrics compared with traditional CI/CD, rule-based recovery and basic AIOps baselines. Failure detection time decreased from 120 seconds to 35 seconds, MTTR de-creased from 25 minutes to 8 minutes, pipeline success rate increased from 78% to 94%, recovery accuracy increased from 68% to 93% and automation efficiency increased from 35% to 88%.Future work will extend the framework to GitHub Actions, GitLab CI and Azure DevOps. Additional improvements in-clude Kubernetes-native operators, Prometheus and Grafana integration, automated pull request generation, explainable AI reports, enterprise-grade security policies and large-scale pro-duction evaluation.

References

[1] J. Humble and D. Farley, Continuous Delivery: Reliable Software Releases through Build, Test, and Deployment Automation. Boston, MA, USA: Addison-Wesley, 2010. [2] G. Kim, J. Humble, P. Debois, J. Willis, and N. Forsgren, The DevOps Handbook. Portland, OR, USA: IT Revolu-tion Press, 2016. [3] B. Beyer, C. Jones, J. Petoff, and N. R. Murphy, Site Reli-ability Engineering. Sebastopol, CA, USA: O’Reilly Me-dia, 2016. [4] Y. Dang, Q. Lin, and P. Huang, “AIOps: Real-world chal-lenges and research innovations,” in Proc. IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice, Montreal, Canada, 2019, pp. 4–5. [5] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly de-tection: A survey,” ACM Computing Surveys, vol. 41, no. 3, pp. 1–58, 2009. [6] R. Bommasani, D. A. Hudson, E. Adeli, et al., “On the opportunities and risks of foundation models,” arXiv preprint arXiv:2108.07258, 2021. [7] J. Wei, X. Wang, D. Schuurmans, et al., “Chain-of-thought prompting elicits reasoning in large language models,” in Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 24824–24837. [8] T. Kojima, S. S. Gu, M. Reid, Y. Matsuo, and Y. Iwa-sawa, “Large language models are zero-shot reasoners,” in Advances in Neural Information Processing Systems, vol. 35, 2022, pp. 22199–22213. [9] R. S. Sutton and A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed. Cambridge, MA, USA: MIT Press, 2018. [10] S. Amershi, M. Cakmak, W. B. Knox, and T. Kulesza, “Power to the people: The role of humans in interactive machine learning,” AI Magazine, vol. 35, no. 4, pp. 105–120, 2014. [11] Jenkins, “Jenkins documentation,” Available: https://www.jenkins.io/doc/, accessed 21 June 2026. [12] Kubernetes, “Kubernetes documentation,” Available: https://kubernetes.io/docs/, accessed 21June 2026. [13] Docker, “Docker documentation,” Available: https:// docs.docker.com/, accessed 21 June 2026.

Copyright

Copyright © 2026 N. Vijay, A. Jitendra. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.