Agent safety moves from models to runtime-level governance

Tushar Jain frames the next agent challenge as safety, not intelligence. He argues that as agents expand tasks dynamically, they cross trust boundaries and require just-in-time, scoped access.

Current permission models are static and insufficient. He proposes a runtime layer that provides containment, scoped access, and intent-based permissions across all models and harnesses. Docker's SPX demo shows sandboxes with network controls, credential injection, and task-specific access.

The demo demonstrates local-to-cloud portability and orchestration across multiple sandboxes. However, the intent-based access system is still early and unproven in production. The claim that runtime solves safety across all models is aspirational.

Engineering leaders should watch how this evolves, but the real constraint is governance policy design, not runtime technology. Teams need to design workflow boundaries before tools can enforce them.