I let an AI agent compose a K8S golden path, then cutover to EKS behind the same domain.

In Part 1, an AI agent put a CDN, a certificate and our own domain in front of Marketing's app. Two days later they came back. Their developer had rebuilt the page as a container: people can now submit their own quotes, and there's a live counter. Which means it keeps state, and the Lambda can't hold it. The platform team had already decided: containers go on Kubernetes, there's no cluster yet, so build it properly as a golden path the next team can reuse. Same for the network. Reuse the encryption key we already have. And the address everyone has already shared does not change. So I gave the ticket to a fresh agent with no context and no AWS keys, and asked it to compose that path: EKS, a VPC, an image registry, a DynamoDB table, a load balancer controller and the workload. Then a second fresh agent moved the CDN from the Lambda to the cluster, behind the same address. This was recorded live on LinkedIn, so there's real waiting while EKS builds. Two things I set up beforehand so the session didn't turn into an hour of watching a progress bar: the modules were already imported into the catalog, and the cluster we cut over to was built that morning from the same request. The path the agent composes on camera is smoke-testing in the background the whole time. It's still running at the end, so it's saved but not published yet. Along the way I ask for m5.large nodes. The platform team only allows t3.small and t3.medium in non-prod, so it's refused, twice, and every refusal lands where the platform team can see it. And because I had time left, we try day two live: upgrade Kubernetes, add Argo CD (which wasn't in the catalog at all), and run two replicas. ⏱ CHAPTERS 0:00 Where Part 1 left off 1:00 The new ticket — keep what people send 1:56 The platform team's rules 3:08 How I'd architect it 4:44 A fresh agent, no credentials 6:26 It jumps straight to the cutover 8:05 Confirming what it understood 9:30 Reuse the key we already have 10:51 Composing a path vs. saving one 12:22 The path review 13:57 Answering the agent's questions 14:41 Refused — m5.large in non-prod 16:38 Every refusal is counted 18:12 The smoke test 21:55 The cluster I built this morning 24:18 Cutover, with a second fresh agent 27:16 Blue/green, while CloudFront updates 28:17 The Terraform behind the path 30:51 Same address, new app 31:37 What a developer sees 32:24 Rollback — and the real point of no return 34:10 Day two: Argo CD, an upgrade, two replicas 39:28 Revision 8 42:52 Who composes the golden path? 43:45 Next: platformising Terraform you already have 🔗 THE PLATFORM Archie — the governed layer between the agent and the cloud: https://askarchie.io I'm looking for a few teams to try it with: read-only first, in your own sandbox, with real tickets instead of my made-up ones. 📺 NEXT EPISODE Can we platformise what already exists? Importing existing Terraform into Archie, breaking it into modules and a golden path, then deploy, drift, upgrade and rollback on top of it. Part 1:    • I asked an AI agent to put a real domain i...   #PlatformEngineering #DevOps #AIAgents #InternalDeveloperPlatform #InfrastructureAsCode