Long-duration agents in sandbox: architecture, security, and task recovery in 2026
An agent who works for hours needs checkpoints, budgets, persistent states, and access limits. The model is only part of the system: the harness and the sandbox determine what it can actually do.