Your agents now live on six platforms. Dataiku and Classie are selling the layer above all of them

Dataiku’s chief executive framed the problem with a line that lands with any bank CIO: ask a bank how many servers it runs and you get an exact answer; ask how many AI agents it runs and you get a guess.

The guess is not carelessness. Agents are now built in Copilot Studio by business teams, in Azure AI Foundry and Amazon Bedrock by engineering, in Salesforce and ServiceNow by application owners, in Databricks and Snowflake by data teams, and in browser extensions by employees who never asked anyone. Each platform has its own console, its own logs and its own idea of what an agent is.

Two announcements in a week target exactly this gap. Dataiku launched Agent Management, which it says reaches general availability in October. Classie launched Supervise, which it says is generally available now. They come at the problem from opposite ends, and the difference is the useful part.

Two products, two entry points

Dataiku starts from the platforms. Agent Management connects to agent platforms through their native APIs and log streams, and to custom environments through OpenTelemetry. Dataiku lists AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio, Azure Foundry, Salesforce Agentforce, Snowflake Cortex and its own platform. From there it builds one inventory with each agent’s structure, tools, models, owner and purpose, tracks usage, cost and quality, and keeps certification, risk and scheduled-test records for high-risk agents. Pricing, per Dataiku, is an annual per-instance license plus metered monitoring per agent.

Classie starts from the edges. Supervise is positioned to discover and monitor both sanctioned and unsanctioned agents across browsers, endpoints and compute environments. Classie says it records runtime transcripts of what an agent did, for whom and with which data, can alert, restrict, reroute or stop actions in real time through an Open Policy Agent engine, attributes token spend to users and agents, and can be deployed privately. Classie has not published detailed product documentation beyond the announcement, so treat the scope as claimed.

Put simply: Dataiku is strongest where agents are built on known platforms; Classie aims at agents nobody registered.

1. Platform APIs and logs
2. OpenTelemetry from custom code
3. Browsers and endpoints
4. One inventory, owners and risk tier
5. Tests, policy and spend

• Sanctioned agents: Platform APIs and logs · OpenTelemetry from custom code
• Shadow agents: Browsers and endpoints

Dataiku's emphasisClassie's emphasis

What a layer above the vendors can do

A cross-platform control plane is good at four things.

Counting. One list of agents, with an owner for each, is the prerequisite for everything else. Federal agencies already work this way for AI use cases: OMB’s memorandum M-25-21 requires them to keep and publish AI use case inventories. Large banks have kept model inventories for more than a decade under SR 11-7, the model risk guidance the banking agencies revised in April with a focus on organizations above $30 billion in assets. Agents are the obvious next line in that kind of register.

Tiering. Not every agent deserves the same scrutiny. A control plane that records tools, data access and autonomy can sort agents into risk tiers and attach the right tests and reviews to each.

Comparing. Cost per agent, quality trends and drift across platforms in one view, instead of six consoles with six definitions.

Finding orphans. Agents whose owner left, whose tests stopped running, or which nobody has called in months. These are the agents that cause incidents.

What it cannot do

It cannot see deeper than the platforms expose. Integration depth varies by framework. A control plane that reads a platform’s API may see an agent’s configuration but not every tool call. Ask, for each platform you run, which fields are actually collected.

It does not replace platform-native controls. Identity, permissions and data access are still enforced where the agent runs. In Microsoft environments, Entra Agent ID gives agents their own identities with sponsors and Conditional Access, and Microsoft is converging its agent registry under Microsoft Agent 365. A cross-platform layer should read from those controls, not compete with them.

Its telemetry is sensitive. Runtime transcripts contain prompts, retrieved documents and outputs. As we argued in our piece on agent observability, the monitoring data needs the same custody decisions as the data it describes.

How to decide

The test is simple: count your agent platforms. If nearly everything runs in one ecosystem, start with that ecosystem’s registry and identity controls and add a cross-platform layer later. If you already run agents on three or more platforms, or if business units buy agents on their own, a layer above the vendors starts to pay off, because no single vendor’s registry will ever list its competitors’ agents with the same care.

That is the logic of owning vendor-neutral layers once, which we laid out in our piece on what to build once across every agent. The inventory is one of those layers.

Before signing, run a two-week test:

1. Reconcile. Compare the tool’s discovered inventory with what each platform team believes it runs. The gaps are your finding.
2. Check depth per platform. For your top three platforms, confirm which events are collected: configuration only, or tool calls and data access too.
3. Assign owners. Every discovered agent gets a named owner within the trial. Agents nobody claims are candidates for shutdown.
4. Decide where transcripts live. Retention, access and region for the control plane’s own data, in writing.
5. Wire one action. Pick one policy, such as blocking an unregistered agent from a sensitive system, and test it end to end.

The bottom line

Agent sprawl is following the path of SaaS sprawl and cloud sprawl: first growth, then a bill, then an inventory. Dataiku and Classie are betting the inventory phase starts now. Whether you buy one of them or build on your main platform’s registry, the first deliverable is the same: a list of every agent, with an owner next to each one. Without that list, the question of who answers for an agent’s actions has no one to address.