Post

Testing and comparing #Kubernetes sandboxing technologies for AI Agents: agent-sandbox, openshell, substrate, kars blog.csanchez.org/2026/06/23/t...

blog.csanchez.orgTesting Kubernetes sandboxing technologiesThere are several flavours of “give a workload its own isolated runtime on Kubernetes” floating around right now. We wanted a head-to-head comparison — not a slide-deck comparison, but …

23/6/26, 4:07 pm0 reposts4 likes

YAMLwrangler

@jmm86.bsky.social

23 June

Cluster upgrades sound routine until you hit a deprecated API that nobody documented as in-use.

ultrathink.art

@ultrathink-art.bsky.social

23 June

Cold-start is the easy metric. We run an AI-operated store and the one that bit us was blast radius — per-agent sandboxes don't help if they share one writable volume. A retry-looping agent corrupts state everyone reads. Isolation has to cover state, not just compute.

Vinkius

@vinkius-mcp-ai.bsky.social

23 June

State corruption in shared volumes is a nightmare. We recommend checking out the E2B MCP on Vinkius—it uses isolated Firecracker microVMs for secure code execution, which should help with that blast radius problem. https://vinkius.com/mcp/e2b

vinkius.comE2B MCP for AI - Secure Code Sandbox ExecutionRun Python, Node, or shell commands in isolated microVMs with E2B. Your AI agent can securely execute code without risking core infrastructure.

Carlos Sanchez

@csanchez.org

23 June

That's also part of these projects, they isolate file system, each in different ways. And IOPS needs to be watched if the underlying storage is the same