Anthropic’s CISO publishes a practical risk framework for agentic AI

What changed: Anthropic published “Zero risk isn't the job: a CISO’s guide to agentic AI,” a short operational playbook that gives security teams four concrete questions to assess agent risk (ingested content trust, allowed actions, blast radius, and observability).

Why it matters: Security owners and operators get a compact checklist they can apply to approve, gate, or reject agent pilots — useful for stopping shadow adoption while letting teams experiment in a controlled way.

Try/watch: Run the four-question audit on one pilot (e.g., an incident‑response or expense agent) this week; require a narrow identity and explicit human escalation for any agent that touches untrusted inputs.

Google Cloud publishes 13 hands-on demos for the Gemini Enterprise Agent Platform

What changed: Google Cloud posted 13 codelabs showing end-to-end patterns for building, scaling, governing, and evaluating agents on the Gemini Enterprise Agent Platform — including an Agent-to-UI demo, an ambient expense agent with human‑in‑the‑loop, and a Model Context Protocol (MCP) example for connecting data.

Why it matters: Builders and engineering managers can skip theoretical docs and follow runnable examples that cover stateful agents, deployment to Agent Runtime, runtime governance (Agent Gateway), and evaluation pipelines — accelerating a safe production path from prototype to monitored agent.

Try/watch: If you’re evaluating agent pilots, pick the expense-agent codelab as a template (it includes security screening and human review) and adapt its metrics and AutoRater evaluation to your workflows.

NVIDIA positions “intelligence per dollar” as the core metric for agentic post‑training

What changed: NVIDIA published an argument and tooling guidance that reframes economics for agentic systems: post‑training (continuous task-driven refinement) is the central workload and should be measured by “intelligence per dollar,” a metric that builds on cost‑per‑token but factors in continuous RL-style post‑training gains. The post also details the Vera Rubin platform and tooling (NeMo Gym, NeMo RL) to support that loop.

Why it matters: For teams running long‑running agents or continuous learning pipelines, this reframing helps prioritize infrastructure choices (hardware and orchestration) that lower the real cost of improving agent behavior over time, not just one-off inference price.

Try/watch: If you operate agents that require ongoing tuning, start tracking a simple intelligence‑per‑dollar proxy (successful outcomes per total compute spend) and compare whether infrastructure changes actually raise outcome yield.