Abhishek Bhardwaj traces the isolation options available for code-running agents, from ordinary processes and containers through gVisor-style boundaries and microVMs. Each choice trades startup speed, compatibility and security against operational complexity.
A useful sandbox also needs persistence. Snapshots and restore points let long-running agents recover state, branch experiments and support reinforcement-learning workloads without rebuilding the environment after every step.
The final layer is orchestration across a fleet: selecting nodes, controlling networking, exposing approved ports and managing lifecycle events at low latency. Bhardwaj's architecture shows why an agent sandbox is a distributed systems product, not merely a container wrapper.
Watch on YouTube
