3 posts
Run untrusted agent-generated code in Kubernetes with a layered sandbox: gVisor runtime isolation, restricted Pod Security, default-deny networking, quotas and admission policy that enforces all of it.
Instrument LLM agents with OpenTelemetry GenAI conventions, send traces and metrics to Grafana, and build dashboards and alerts for cost, latency, guardrail blocks and tool failures across AWS Bedrock, GCP Vertex AI and Microsoft Foundry.
Run MCP servers as isolated containers behind the Docker MCP Gateway, keep secrets out of config files, harden your own tool servers, and build an agent in Compose that can only reach the tools you approved.