How many Flux agents should a large Kubernetes cluster use?

One Flux installation per cluster is generally a good default. Add another only when you need a genuine isolation boundary, such as different credentials, tenant administration, upgrade ownership, or a smaller blast radius. Separate repositories alone usually aren’t a strong reason to run separate agents. You can also use dependsOn between Kustomizations when bootstrap order matters. Otherwise, keep the topology simple and let observed API-server load or reconciliation delays guide any future split.

I wouldn’t install Flux separately for every microservice. Repository ownership and controller topology are separate concerns, so a single cluster-wide installation can still support multiple teams and projects through GitRepository and Kustomization resources. The important scaling metric is reconciliation workload rather than raw service count. Monitor controller CPU and memory, reconciliation latency, Kubernetes API throttling, and source-fetch failures. If those become problematic, shard the controllers and keep related sources and Kustomizations assigned to the same shard. With several thousand services and no symptoms today, your current design sounds reasonable.

Using the bootstrap command through infrastructure provisioning is perfectly valid. An operator may offer a different management workflow, but it isn’t automatically more efficient or required for this architecture. If the current installation is stable and Terraform handles the initial cluster setup reliably, there’s no obvious need to change it just because the cluster manages many services.