The run of briefs before this one tracked authorization as a boundary the card networks and regulators were slowly co-signing. The question was who gets to approve an agent's action. Today the question moves one layer down: once an agent is running, who can see it, interrupt it, and bound what it writes to disk.
Three signals point at the same answer, and it is infrastructure, not policy. OpenAI's Control Room ships a central registry of every model workload with hard-stop authority over any running job. Snowflake's Cortex AI Gateway puts the same interrupt-and-audit logic at the data layer, so agents are governed at the point of data access and the blast radius shrinks when one misbehaves. Cashfree's Checkout for Agents draws the line at the authorization boundary: agents read and compose, humans authenticate the payment. Three vendors, three layers, one pattern. A control plane that can inspect and terminate non-human work is becoming a shipped feature rather than a governance aspiration.
The Claude Code credential leak is the counterweight, and it is the one to sit with. Thousands of settings files carrying live secrets pushed to public GitHub, no attacker action required, through a standard developer workflow. The agent's write scope extended past code into configuration, and nobody had drawn a boundary around what it could touch. This is the failure mode the control-plane vendors are selling against, and it happened at the one layer none of today's gateways cover: the developer's local disk. The 2026-W43 index puts Governance & Ethics at 79, the top dimension in its category. That score measures intent. The Claude Code incident measures operation, and the gap between the two is exactly where enterprise risk now lives.
So the move this week is to stop treating agent governance as a document and start treating it as a running system with two properties: a registry of what is executing, and an interrupt authority over it. Control Room is the reference architecture for the first. The second question is harder, because it is organizational, not technical. Who, in your firm, is authorized to kill a production agent at 3am, and do they have the switch. If the answer is a policy PDF and not a named person with a button, you have the Claude Code posture, not the Control Room posture.
The procurement angle sharpens this. ChatGPT Enterprise is now buyable through the AWS Marketplace against committed cloud spend, which means AI tools enter the firm before the governance team is in the room. Map which agents are arriving through marketplace agreements, because those are the ones most likely to run ungoverned.
Watch item: the Secure Technology Alliance and Merchant Advisory Group working group on agentic commerce standards. Its focus on agent authentication and false decline rates puts non-human buyer identity on the payments roadmap. Track it for published frameworks. The Paypers coverage.
Index Reference · Applied AI Index 2026-W43, the latest published issue (Sunday, October 25, 2026)
WatchThe Secure Technology Alliance and Merchant Advisory Group partnership on agentic commerce standards is early-stage, but the focus on agent authentication, false decline rates, and payment card selection puts identity governance for non-human buyers on the payments industry roadmap. The Paypers coverage is worth tracking for published frameworks or recommended practices from this working group.