An Alliance Is Not a Control Plane

· 6 min read agentsgovernanceservicenow

Navy and amber hero graphic with bold text: An Alliance Is Not a Control Plane

On October 28, ServiceNow announced the Control Tower Alliance, a partner program built around its Super Intelligence Control Tower. Unite.AI’s summary of the announcement lists eleven founding members: Anthropic, Amazon Web Services, Cisco, HiddenLayer, Kong, NVIDIA, Okta, OpenAI, Stripe, WitnessAI, and Zenity. ServiceNow describes the tower as a centralized control plane to discover, secure, govern, observe, and measure enterprise AI. Partners build solutions that run on the tower, submit them to ServiceNow qualification and certification against shared standards for agent identity and safety guardrails, and sell jointly.

Give the announcement its due first. Amit Zavery, ServiceNow’s president, chief product officer, and chief operating officer, said that agents have multiplied to the point where no single solution can see or protect everything on its own. That sentence is the most honest one in the release. It is also the one that limits the claim in the headline. If no single vendor can see the whole estate, then a central tower does not govern the estate by itself. It governs what its partners feed it and what its partners enforce. The alliance is the admission. What you do with the admission is the work.

I have made a version of this point before. In The Blocker Is Not the Control Plane, I argued that one runtime check is one control among several, and that a demo of the check tells you little about the system around it. The same applies one level up. A roster is not a control. Certification is not a control. An inventory feed is not a control. Each of those can support a control, and confusing the support for the control is how a shop ends up well documented and unprotected at the same time.

Kong’s own announcement states the dividing line in one sentence. Augusto Marietti, Kong’s CEO and co-founder, said: “AI governance depends on enforcing what agents can access and do.” Kong’s part of the alliance is an API gateway that sits in the path of model, tool, and agent traffic and can apply policy as the traffic passes. That is a different kind of thing from a partner logo. One sits in the path and can say no in real time. The other sits in a program and says the members agreed to work together. Both can be worth having. Only one of them stops an action.

The Cisco membership shows the other half of a working arrangement. In Cisco’s write-up of its own membership, the company describes two integrations that already exist on the ServiceNow Store. The first brings the AI assets Cisco AI Defense discovers into the Control Tower inventory. The second sends security detections into ServiceNow AI Security Exposure Management, where teams assign owners and track each fix. Cisco states the split plainly: Cisco remains the source of security insight and protection, and ServiceNow serves as the system of action for governance and remediation. Read that split carefully, because it is the architecture. Detection lives with the tool that can see. The record, the owner, and the fix live in the system where work gets assigned and closed. Neither half is the whole control. Together they can be one.

That is the pattern I would generalize, and it cuts against the word centralized in the announcement. Centralize the record: the inventory, the policy decision, the owner, the evidence, the fix status. Federate the enforcement: the gateway, the identity provider, the runtime, the tool that sits where the agent actually acts. My estate does not run in one vendor’s cloud, and neither does yours. Agents act in a dozen places. A control that only works inside one platform governs the part of the estate that needed it least, because that part was already inside a managed boundary.

Four tests sort an alliance integration into what it actually is before you count it in a governance review.

First test: in path or in report. Ask whether the integration can block an action, revoke a credential, or stop an agent while work is running, or whether it sends findings and inventory records to the tower after the fact. The Cisco asset feed is in report. The Kong gateway claim is in path. A report feed is useful. It is not enforcement, and a review that scores it as enforcement overstates the posture by exactly the gap an incident will later measure.

Second test: certified against what, and for how long. Qualification and certification happen at listing time, against a version of a product and a set of shared standards. Agents, models, and tool connections change after listing, sometimes weekly. I covered that drift problem in The Agent You Approved Is Not the Agent Running. Ask what change on the partner side voids the certification, who retests, and how you would know. A certification with no stated expiry or retest trigger is a photograph, not a control.

Third test: what happens when the tower cannot be reached. If the central plane is down, partitioned, or simply not deployed in an environment, does local enforcement keep working on the last policy it received, or does control stop with the connection? Enforcement points need to hold their policy locally and keep applying it while disconnected, with the record catching up when the connection returns. A design where the enforcement point asks the tower for permission on every action has built a single point of failure and called it governance.

Fourth test: who closes the loop. A detection with no owner is a notification. A finding with an owner and a tracked fix, the pattern in the Cisco exposure management integration, is a control loop. Ask of every alliance feed: where does this land, who is named on it, what state counts as closed, and what evidence proves the fix. If the demo ends at a dashboard, the loop is open.

None of this makes the alliance a bad buy. A curated set of integrations that have passed qualification testing saves real procurement time, and a common record for AI assets is work every enterprise has to do somewhere. I made the cross-platform version of this argument yesterday in Govern the Agents You Did Not Build: the agents you did not build still need identity, containment, and a record. The alliance is one credible way to assemble those parts. Just score it as parts. Count the controls at the point of action, where an agent can be blocked, revoked, or shut down, and count the record where an owner closes a fix. Count nothing for the roster itself.