Architecture AI Updates: October 2, 2026

1. Sandboxes, Canaries and Autonomy Levels for Agents in Production Operations

InfoQ. An InfoQ panel with Michael Hausenblas (Genesys SRE), Sujana Sooreddy (Netflix), and Noam Levi (Groundcover) lays out patterns for letting agents into production operations: develop agentic systems sandbox-first, treat security as a separate layer rather than something baked in, and keep every release behind canary promotion with automated rollback. The panelists argue that high-cardinality observability data makes agent diagnoses more precise than coarse metrics, and that agents should get structured topology, metadata, and decision records rather than raw payloads, with zero-data-retention boundaries by default. They also recommend capturing agent trajectories with confidence scores before execution and calibrating each task’s decision authority with an L0 to L4 autonomy scale like Google SRE’s. Source