🔥 Top Story
AI adoption isn’t the same as AI usage
Engineering organizations frequently rely on usage metrics such as seat activations, token consumption, and code completion rates to evaluate AI adoption. However, these vanity metrics often mask a gap between superficial tool interaction and meaningful workflow transformation. Following Goodhart's Law, targeting usage numbers incentivizes developers to optimize for metrics without improving overall software quality. Genuine adoption requires shifting from ephemeral individual habits to durable team norms where recurring manual tasks are permanently offloaded. For instance, Webflow achieved sustainable adoption by managing AI prompts as version-controlled code artifacts within shared repositories. Ultimately, scaling agentic workflows depends on addressing human code-review bottlenecks and establishing domain-specific evaluation rubrics rather than trusting generic benchmarks.
💡 Why it matters: When integrating AI into cloud and DevOps infrastructure management, teams must focus on version-controlling prompt artifacts and establishing evaluation rubrics rather than tracking raw usage metrics to achieve durable operational efficiency.
🔗 Read more · The New Stack
DevOps & Infrastructure
Why your KubeVirt VMs can’t move between clusters — and how EVPN fixes it
The New Stack
Live-migrating KubeVirt virtual machines across Kubernetes clusters requires preserving original IP and MAC addresses to prevent session disconnections in stateful applications. Traditional cross-cluster VM migrations rely on stretched Layer 2 VLANs, which involve manual network switch reconfigurations and prolonged ticket approval cycles. OpenPERouter resolves this network bottleneck by expressing EVPN/VXLAN overlays as native Kubernetes Custom Resource Definitions such as Underlay, L2VNI, and L3VNI. By deploying identical L2VNI specifications across clusters, VMs achieve seamless Layer 2 adjacency over existing IP underlay networks. Separating migration traffic into a dedicated VXLAN segment isolates heavy memory-state transfers from regular application workload traffic. Coupled with IPAM solutions like Whereabouts and KubeVirt migration CRDs, platform teams can execute multi-cluster VM mobility declaratively without physical infrastructure dependencies.
💡 Abstracting EVPN/VXLAN overlays into Kubernetes CRDs enables platform teams to execute multi-cluster VM live migrations declaratively without physical network reconfigurations, greatly enhancing disaster recovery and workload rebalancing agility.
Five AI rivals just backed a shared plugin standard. Here’s why it matters for developers.
The New Stack
Major AI industry leaders including OpenAI, AWS, Cursor, GitHub, and Microsoft have endorsed the new Agent Plugins 1.0.0 standard initiated by Vercel. Agent Plugins introduces an open, vendor-neutral package format that bundles AI Agent Skills and Model Context Protocol (MCP) servers using a standard plugin.json manifest. This standard eliminates fragmentation across diverse agent clients like ChatGPT, Cursor, GitHub Copilot, and VS Code by establishing unified discovery and loading mechanics. By decoupling plugin definitions from vendor-specific runtimes, developers can write reusable AI extensions once and deploy them across compatible client tools. Industry experts highlight that while this open specification reduces integration friction, it places greater responsibility on individual clients to handle governance and security permissions. The initial 1.0.0 release focuses specifically on packaging Agent Skills and MCP servers, with commands and execution hooks reserved for future specification updates.
💡 While standardizing AI agent plugins significantly improves ecosystem portability, platform engineers must enforce granular client-side access controls and sandboxing to prevent security compromises across distributed environments.
This digest was collected from RSS feeds and summarized by AI (Claude). See the original links for full details.