Mendrel advocates for running AI agent harnesses outside sandboxed environments rather than inside them, particularly for multi-user enterprise deployments. This architecture keeps credentials secure, enables efficient resource utilization through sandbox suspension, and supports shared skills and memories across teams—but requires solving three critical infrastructure challenges: durable execution (using workflow orchestration like Inngest), sandbox cold-start optimization (via Blaxel for 25ms resume times), and distributed filesystem abstraction for agent state management. For IT organizations, this represents a fundamental shift in how to architect AI agents for production, requiring investment in orchestration layers and state management systems rather than relying on simple containerized approaches.
Mendrel advocates for running AI agent harnesses outside sandboxed environments rather than inside them, particularly for multi-user enterprise deployments. This architecture keeps credentials secure, enables efficient resource utilization through sandbox suspension, and supports shared skills and memories across teams—but requires solving three critical infrastructure challenges: durable execution (using workflow orchestration like Inngest), sandbox cold-start optimization (via Blaxel for 25ms resume times), and distributed filesystem abstraction for agent state management. For IT organizations, this represents a fundamental shift in how to architect AI agents for production, requiring investment in orchestration layers and state management systems rather than relying on simple containerized approaches.