Agent Sprawl Becomes a Product Category: Dataiku, Broadcom and the Governance Land Grab

TL;DR

Dataiku unveiled a standalone Agent Management product on 24 September, generally available in October 2026, that inventories, scores and audits agents across six enterprise platforms plus its own stack, with OpenTelemetry covering everything else.
It landed inside a 14-day window that also produced Broadcom’s AgentMinder, Okta’s Agent SSO and IBM’s watsonx Orchestrate AgentOps agent — four vendors, four different infrastructure layers, no shared standard.
The commercial trigger is measurable: an independent Omdia study for Cisco found 51% of large enterprises already run agentic AI acting in production network operations, while 95% say non-agentic AIOps tooling falls short.
The category currently sells inventory. What enterprises still lack is pre-authorisation and a certification trail — an inventory is not a control plane.

Something changed in enterprise software last month, and it was not a model release. It was the sudden commercial viability of the least glamorous question in the agent stack: how many of these things are running inside my company, and who authorised them?

The 14-day window that created a category

Dataiku unveiled a standalone product called Agent Management at its annual conference on 24 September 2026, with general availability set for October. The pitch is deliberately unsexy: discover and monitor every AI agent across platforms in one place, track quality and cost per agent, and control risk from a single console. (Source : Dataiku — Agent Management)

The connective tissue matters more than the feature list. Agent Management connects to AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio and Azure AI Foundry, Salesforce Agentforce and Snowflake Cortex, plus Dataiku’s own platform — with OpenTelemetry support for anything custom. That last detail is the tell. A governance product that ships both native connectors and an open telemetry path is conceding that the agent estate will never be single-vendor. (Source : AI Magazine — Dataiku: Solving AI Sprawl and Risk with Agent Management)

It was not alone. Broadcom unveiled AgentMinder at VMware Explore on 31 August, positioning it as a way for enterprises to use AI agents while ensuring they follow company rules and security standards — Broadcom says it uses the product internally to scale its own agentic estate. (Source : Broadcom — Broadcom Unveils AgentMinder)

The clustering is the story. Between 24 August and early September, four vendors shipped standalone agent-governance products from four different layers of the stack: Okta with Agent SSO (identity), IBM with watsonx Orchestrate’s AgentOps agent (orchestration), Broadcom with AgentMinder (security) and Dataiku (observability). Each looked at the same buyer problem and assumed their own layer was the natural place to solve it. (Source : Yahoo Tech — Dataiku Ships Standalone Agent Governance)

The numbers that made vendors move

Enterprise AI governance has been discussed for two years. What changed is the denominator. An independent study conducted by Omdia for Cisco, published on 23 September 2026, surveyed 1,000 IT and network operations leaders at organisations with 500 or more employees. The headline finding is not a projection: 51% already run agentic AI that acts in production network operations, in a model the study calls AgenticOps, where humans set direction and guardrails while agents sense, reason and act across domains. (Source : Cisco Newsroom — AgenticOps Scaling Quickly in the Enterprise)

Two supporting figures explain the urgency. 84% of respondents expect to reach an AI-led operating model within 12 months, and 95% say their existing, non-agentic AIOps tools fall short of what that requires. The study also quantified the noise floor: organisations generate roughly 4,100 monitoring alerts per day on average, which is precisely the volume that makes human triage structurally impossible and rule-based automation insufficient. (Source : StockTitan — Cisco AI Study: 51% Run Agentic AI in Production)

Read against each other, the Cisco figure and the Dataiku launch describe the same gap from opposite ends. Half of large enterprises have already delegated production actions to software that reasons, and the tooling that was supposed to supervise that delegation was designed for systems that do not. That is a category opportunity with a hard deadline.

The accountability gap is the product

The most useful framing comes from Dataiku’s own positioning: a company may have thousands of agents built over months by different departments, while oversight, risk definition and measured business outcomes exist for a small fraction of them. The majority of the broader agent portfolio runs without clear accountability. (Source : Technology Magazine — How Dataiku Solves Enterprise AI Agent Sprawl and Risk)

That description is worth taking seriously as an inventory problem, because it is one an inventory can actually answer. If an agent exists, it consumed a credential, made a network call and cost tokens — all of which are observable, and none of which require the agent’s author to cooperate. Discovery by telemetry is the only approach that survives contact with real enterprise sprawl, where the people who built last year’s agents have often left the team.

The structural constraint is that governors and builders come from different disciplines. Identity vendors see the problem as credential lifecycle. Orchestration vendors see it as workflow. Security vendors see it as runtime policy. Observability vendors see it as telemetry. All four are correct and none is sufficient, which is why the next twelve months of this category will be fought over who owns the agent record of truth rather than over features.

Broadcom’s approach is instructive on this front. Its deny-by-default agent runtime work, shipped in the same quarter, treats agent permissions as infrastructure policy rather than application logic — the agent does not get a broad role and a promise, it gets a scoped identity with a default of no access. (Source : The Agent Report — Broadcom, VMware and the Deny-by-Default Agent Runtime)

What is still missing

Three capabilities are absent from every product shipped in the September window.

Pre-authorisation. Inventory tells you an agent exists. It does not tell you whether this agent, at this hour, is permitted to call this external endpoint with this budget. That requires a policy decision point in the request path, which is a latency-sensitive place to put governance and therefore the hardest part of the problem.

Certification. An agent registry is only as useful as its ability to distinguish a tested agent from a prototype someone forgot to turn off. Nothing shipped this month issues a scoped, expiring declaration of what an agent has been validated to do — which is why the runtime posture (deny by default) is currently doing the work that certification should.

An audit trail that a regulator accepts. Cost per agent and quality scores are operational metrics. An incident review needs permission scope, egress path, model version, and the human who approved the deployment. That record exists in fragments across four vendors and is nobody’s deliverable.

The demand signal is unambiguous: 51% of large enterprises are already past the point where agent governance is hypothetical work. The supply side has responded with inventory, which is the correct first product and the wrong last one. The vendor that converts an agent census into an enforceable permission model owns the next layer of enterprise AI — and given how many well-funded companies just entered the same 14-day race, it will not be obvious for a while who that is. Agent platforms are also rushing to make agents buildable by non-engineers, which will make the census harder to keep accurate: work management vendors now ship drag-and-drop agent builders to the same business units that will show up in the next audit. (Source : The Agent Report — monday.com AI Agent Builder)

FAQ

Why did agent governance products all launch in the same fortnight?

Three reasons converged in September 2026. Enterprises crossed a measurable threshold (51% of large firms running agentic AI in production, according to Omdia’s study for Cisco), identity and observability vendors had finished shipping the primitives needed to attach a registry to real telemetry, and the natural enterprise budget cycle for a new control-plane category starts in Q4.

What does Dataiku Agent Management actually connect to?

Six enterprise platforms plus Dataiku itself: AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio and Azure AI Foundry, Salesforce Agentforce and Snowflake Cortex. Custom environments are covered through OpenTelemetry support.

Is an agent inventory enough for compliance?

No. An inventory answers what exists. Compliance questions usually ask what an agent was permitted to do, under whose authority, at the time it acted. That requires policy enforcement in the request path and an immutable audit record, neither of which is solved by discovery alone.

How do these products differ from existing AIOps tooling?

Existing AIOps tooling was built to correlate alerts from infrastructure that does not reason about its own goals. The Omdia/Cisco data makes the distinction blunt: 95% of respondents say their non-agentic AIOps tools fall short for agent-driven operations, and organisations average about 4,100 monitoring alerts daily, which is beyond manual triage.

What should an enterprise do before buying?

Run the census itself. Agent sprawl is discoverable through telemetry — network egress, credential use, token spend — without needing cooperation from the teams that built the agents. Whatever governance product gets bought, that discovery pass becomes its baseline dataset and the first honest measure of how much of the estate is actually known.