XR's capability-based security model

How we think about agent security, from sandboxing to human-in-the-loop confirmations, signed skills, and audit trails.

XR 3.1 G represents a culmination of lessons learned from millions of agent runs, thousands of developer feedback sessions, and deep investment in a coherent runtime architecture rather than a patchwork of wrappers.

What’s new

We’ve rewritten the core in Rust to drive cold starts below 10ms, redesigned the skill graph to make composition first-class, and shipped a model router that understands cost, latency, and capability tradeoffs.

What’s next

Expect faster local models, richer multi-agent orchestration, and deeper editor integrations over the coming year. We’re just getting started.