Agent Substrate: Sandboxing is Not Enough

Sandboxing is necessary. It isn't sufficient.

Most AI agents sit idle 99% of the time, yet on Kubernetes they still run as always-on pods. Agent Substrate changes the model: agents become actors that run on a shared worker pool for short bursts, suspend when idle, and resume in milliseconds with their memory and files intact.

In this hands-on workshop, Lin Sun and Christian Posta install Agent Substrate on a Kind cluster, walk through its architecture, and run five labs covering suspend and resume, persistent sandboxes, shared worker pools, request parking, and egress policy.

• Pack far more agents per clusterPods become execution workers, not the deployment unit, so idle agents stop holding CPU and memory.
• Keep state across suspend and resumeGolden snapshots and full-state checkpoints bring an agent back with its memory and workspace exactly as it left them.
• Give every agent its own identityEach actor carries a durable SPIFFE identity that follows it from worker to worker, separate from the pod it runs in.
• Enforce policy on the way outOutbound calls carry the actor's identity to an agentgateway egress point that can allow, block, or inject credentials.

Workshop chapters

• 02:15Why agents shouldn't run as always-on pods, and where kagent fits
• 09:00Installing Agent Substrate with agentgateway as the router
• 09:40Architecture: control plane, worker pools, and snapshot storage
• 11:30Agent identity: actor templates, actors, and SPIFFE
• 16:10How requests and outbound calls flow through Substrate
• 21:10Lab 1: Suspend and resume with state preserved
• 27:50Lab 2: A sandbox whose files survive suspension
• 33:30Lab 3: Multiple actor templates on one shared worker pool
• 40:50Lab 4: Request parking when every worker is busy
• 50:30Lab 5: Egress policy and actor identity at the gateway

Go deeper