How can we make use of AI agents with existing GitLab CI/CD pipelines?

We currently utilize GitLab CI/CD pipelines to deploy our Kubernetes clusters. I am planning to integrate AI agents to gather cluster-related data and automate the deployment process. I am considering two primary approaches:

Replacing GitLab CI/CD with Agentic Workflows: Is this a reliable strategy? With GitLab, we can easily rerun a pipeline from the point of failure after fixing an issue. I am concerned that agentic workflows might lack this granular state management and retry capability.

Invoking GitLab CI/CD via Agents: In this scenario, the agent would trigger the GitLab API to start a pipeline or job. However, this seems to add more complexity than simply triggering them manually. How can I implement this efficiently?

Ultimately, my goal is to use agents to eliminate manual pre-pipeline preparation and to have them autonomously analyze and fix errors when a pipeline fails. Has anyone successfully implemented AI agents for this purpose?