MCP Tunnels: How OpenAI and Anthropic Bridge Cloud AI to Private InfrastructureBoth OpenAI and Anthropic shipped MCP tunnels in May 2026 to connect cloud-hosted AI products to private MCP servers without exposing inbound ports. A comparison of the architectures, security models, and trade-offs.
Anycast vs Geo DNS: Routing Users to the Nearest PoPA comparison of Anycast and Geo DNS for directing traffic to the closest point of presence, including trade-offs, failover behaviour, and when to combine both.
Link: Claude Platform on AWSAnthropic's native Claude Platform is now available through your existing AWS account — same-day feature availability, no separate vendor relationship.
Instant Agents: How Sub-Second Infrastructure Changes What Agents Can BeA deep dive into Modal's container startup stack — FUSE filesystems, gVisor, memory snapshots, GPU checkpoint/restore — and why instant infrastructure enables a new model of ephemeral AI agents.
Link: Agent Substrate — agentic infrastructure on KubernetesAgent Substrate multiplexes large numbers of stateful agent sessions onto a small pool of Kubernetes Pods, achieving heavy oversubscription with sub-second activation.
Link: GitHub Agent HQ — agents wherever you already workGitHub's Agent HQ unifies AI coding agents across Slack, VS Code, the terminal, and github.com into a single mission control.
Monorepo vs. Multi-repo for AI Coding AgentsAI agents amplify your existing architectural decisions. Good boundaries help agents as much as they help human teams — and bad ones cause the same kind of chaos, just faster.
Bounding your agentAgents have no natural stopping point. Here's how to add one — tracking token spend, wall-clock time, and what to do when a limit is hit.
kro: Kubernetes Operators Without the Codekro generates a CRD and a live reconciliation controller from a YAML ResourceGraphDefinition. Here is how it compares to Helm and proper operators, and why it works surprisingly well as a toy operator framework.
AI agent tooling: Monty and BeadsTwo projects tackling unsolved problems in coding agents: safe code execution and persistent task memory.
Networking AI Agents on Kubernetes (3): Agents Inside vs. Outside the ClusterLLM providers are building agentic APIs that run the reasoning loop externally, inverting the traffic model. This post covers what that shift means for K8s-hosted tools: how to declare an agent portably and what an MCP Gateway actually needs to do.
Networking AI Agents on Kubernetes (2): Local LLM InferenceRunning inference locally in Kubernetes is harder than deploying any other workload. GPU resource model, model storage, and scaling all require different thinking.
Networking AI Agents on Kubernetes (1)AI agents invert the usual traffic model: their traffic is mostly outbound, long-lived, and body-driven. Standard Kubernetes networking primitives don't handle this well.
The Next Programming Language: Formalized HumanSoftware engineers have always used formal languages. As we shift to communicating with AI coding agents, we're ascending to a new level—human language with professional formality, like law or medicine.
Restricted Execution for LLM Agents: Sandboxing Untrusted CodeExploring sandboxing techniques for safely executing LLM-generated code using firejail, sandbox-exec, and go-restricted-runner