From always-on agents to runtime guardrails

OpenAI unveils Dots, always-on AI agents with dedicated cloud computers. Dots brings proactive agents into ChatGPT across web, mobile, and desktop, backed by dedicated cloud computers. Always-on agents make coordination, permissions, and clear task boundaries operational requirements, not edge cases.

Pi Adds MCP Support Through a JavaScript Sandbox. Pi adds MCP with sandboxed execution and deferred tool discovery to keep tools composable without loading everything into context. That’s a practical pattern for building agents that can reach more capabilities while preserving context and execution control.

Cloudflare Containers, rebuilt to scale agent sandboxes. Cloudflare says its redesigned containers start more than six times faster and support runtime configuration, snapshots, and restores. Fast, reproducible sandboxes make it easier to give agents isolated workspaces and recoverable execution—Principle 07 in practice.

Validating AI Models and Agents with Property-Based Testing. Property-based testing checks invariants across generated, meaning-preserving inputs rather than relying only on fixed expected outputs. It gives teams a way to expose inconsistent agent behavior and validate outcomes across a wider range of cases—Principle 16.

Okta moves inline to police what AI agents actually do. Okta’s runtime gateway puts authorization in the action path so organizations can govern or block agent actions after access is granted. That distinction matters: outcome systems need controls over what agents do, not just who or what they can reach.