Really interesting breakdown of the shift from synchronous chat-based agents to background, task-oriented workflows. The approval checkpoints and persistent task state are especially relevant for security-sensitive use cases, where letting an agent operate completely unattended isn’t always ideal.

From the perspective of work we’ve seen at Futurism Security, having clear control points, auditability, and predictable boundaries around what an AI agent can access becomes increasingly important as these systems move from experiments into production. The combination of async execution with human approval could be a practical pattern for handling that balance.