Recently, I ran into a niche problem in the network configuration of a Kubernetes cluster. Because all resources are deployed via GitOps with Argo CD, I could start Codex directly in the GitOps repository, describe the problem, and have it inspect the actual state of the cluster using the Kubernetes CLI. And that is exactly the exciting part:
Codex could compare the configuration in the repository with the live state in the cluster, narrow down the problem, and in one case even perform additional checks using a test container. In the end, it suggested the relevant changes directly in the IaC code.
Without Infrastructure as Code, this would hardly have been possible. Neither I nor the AI would have had a reliable basis for properly understanding what was actually deployed and where the problem was. With IaC, changes are traceable, verifiable, and straightforward to roll out.
To me, it is clear: GitOps, IaC, and AI agents complement each other extremely well.
Why GitOps Is Especially Valuable for Coding Agents
- Git represents the verifiable desired state.
- Changes can be reviewed as a diff.
- Pull requests and CI checks remain control points.
- An agent does not need to make uncontrolled changes directly in the cluster.
- Repository and live state can be compared deliberately.
Necessary Boundaries Still Apply
- Read-only analysis by default.
- Minimum necessary permissions.
- No production changes without review.
- Clearly scoped tools and namespaces.
- Auditability of all changes.
- No secret credentials in the prompt or repository.
A robust GitOps model is useful for more than traditional delivery. It also provides the controllable foundation that coding agents need to analyze technical relationships. In the GitOps Delivery Blueprint, I work with you to develop a clear target architecture for repositories, CI/CD, Argo CD, and controlled changes.
The overview of AI-Assisted Software Development shows how architecture, testing, and delivery can be considered together. The Coding-Agent Readiness Assessment is initially being validated through a limited number of pilot discussions.