Blog
Insights, analysis, and updates from the AI agent economy. Browse by tag · Browse the archive.
Carbon-Aware Kubernetes Scheduling Works in the Lab — Here's What It Takes on Rented Hardware
GreenK8s cut scheduler energy use 39% with its own solar array and Google shifts jobs across continents — a single-region fleet on rented racks can only shift in time. This post maps the published results to their lab conditions and lays out the four-step version worth building: meter with Kepler, cut idle with kube-green, gate one batch queue on a carbon forecast, and write the tenant contract.
Hetzner Plus Scaleway: What a Second Cluster API Provider Really Costs a Self-Hosted PaaS
Scaleway's CAPS gives a Cluster API fleet a second EU-native target with hourly H100-class GPUs Hetzner doesn't sell. Here is the verdict table, the both-sides breakeven math near 48% utilization, and the five ops line items a second provider adds before you adopt it.
Argo CD vs Flux in 2026: 23,100 Stars vs 8,180, and What the Split Means for Your Cluster Fleet
Argo CD leads Flux 23,100 stars to 8,180, with roughly 60% vs 11% GitOps share in 2026 analyses. The gap is architecture — centralized console versus decentralized controllers — tabled here with a decision rule for Cluster-API fleets.
Argo CD 3.3 PreDelete Hooks: Stop Tenant Teardowns From Leaving Orphaned Wrecks
Deleting an app-of-apps tree can orphan PVCs, DNS records, and whole namespaces — or deadlock for 45 minutes. Argo CD 3.3's PreDelete hooks gate teardown on a Job that must succeed first. Tabled here: the three failure modes, a working hook, four gotchas, and a Flux fit check.
88% of AI Agent Pilots Never Reach Production — the Gate Is Deployment Infrastructure, Not the Model
IDC puts the enterprise AI pilot failure rate at 88% and blames deployment readiness, not models. A control-by-control gap analysis of SSO, audit logging, secret scanning, PR gates, license policy, sandboxing, and runbooks against a Render-compatible git-push PaaS with MCP.
Woodpecker CI's 130MB Control Plane vs a 7GB GitHub Runner: What Your PaaS Build Step Really Costs on a €3.79 Hetzner Box
A Woodpecker server and agent idle at about 130MB of RAM while GitHub's standard runner is a 2-core, 7GB VM. We budget both stacks megabyte-by-megabyte on a €3.79 Hetzner CX22 and work out when owning your build fleet beats paying per-minute Actions overage.
Boring Infrastructure on Purpose: What Uncloud's Control-Plane-Free Clustering Keeps — and What It Quietly Gives Up
Uncloud clusters Docker hosts across cloud VMs and bare metal with no control plane and no quorum — 5,200+ stars and a university moving 300+ sites onto it. A concrete keeps-versus-gives-up trade-off table, a dead-machine walkthrough, and why the verdict flips once AI agents operate the fleet.
No SSH? No Problem: Talos Linux's talosctl debug Is a Privileged Escape Hatch That Cleans Up After Itself
Talos v1.13's talosctl debug runs a privileged container in the host's own namespaces — with registry-pull and offline tarball-push image paths — and deletes it on exit. We walk through a real pwru diagnosis and the audit rules for using it safely.
Self-Hosted vLLM on Kubernetes: What Owning Your Inference Layer Actually Costs
CNCF's July 2026 vLLM walkthrough builds self-hosted inference in three Kubernetes resources. A worked breakeven — 10 to 15M tokens a day against Sonnet 5 on a €1,199/month GPU box — plus the four production gaps the lab skips and the five cases where owning the layer wins.
Subscribe
New posts land in your reader as soon as they publish. Pick a format — all three carry the same posts.
Current feeds keep roughly two days of posts so daily polling does not miss a burst. Older entries stay reachable from the feed's next-page link in readers that follow it, or from the blog archive.
Following one topic instead? Browse tags