The Agent Governance Stack Is Forming: How Okta, IBM, Broadcom, and Dataiku Are Building the Missing Enterprise Control Layer (2026)

Between August 24 and September 24, 2026, four enterprise software vendors — Okta, IBM, Broadcom, and Dataiku — shipped standalone AI agent governance products. The window between the first launch and the last was exactly one month. None of them coordinated. All of them identified the same gap: enterprises are deploying AI agents faster than they can govern them, and the governance layer that sits between the agent and the enterprise resource simply did not exist as a product category twelve months ago.

This is not a minor product-release roundup. This is the formation of a new infrastructure layer — the agent governance stack — and every Agentic AI Architect, RPA lead, and automation program manager needs to understand what it means for their 2027 planning.

In this guide, we break down each product, map them to a four-layer governance model, compare their approaches head-to-head, and give you a concrete framework for deciding which layers your organization needs first.

Table of Contents

• Why the Agent Governance Stack Is Forming Now
• The Four-Layer Agent Governance Model
• Layer 1: Identity — Okta Agent SSO
• Layer 2: Orchestration — IBM watsonx Orchestrate AgentOps Agent
• Layer 3: Security and Runtime — Broadcom AgentMinder
• Layer 4: Observability — Dataiku Agent Management
• Bonus Layer: StackGen’s Autonomous Operations Factory
• Head-to-Head Comparison Table
• What Gartner Says: The Guardian Agent Category
• What This Means for RPA Teams and Automation CoEs
• Decision Framework: Which Governance Layers You Need First
• FAQs
• Key Takeaways
• References

Why the Agent Governance Stack Is Forming Now

The numbers tell the story. Gartner projects that 40% of enterprise applications will embed task-specific AI agents by end of 2026, up from fewer than 5% in 2025. Okta’s 2026 “AI Agents at Work” report found that only 34% of organizations apply the same security controls to AI agents as they do to human workers. A Cloud Security Alliance survey of 235 large-enterprise CISOs and CIOs found that 92% lack full visibility into their AI agent identities, and 95% doubt they could detect or contain a compromised agent.

Meanwhile, Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents due to governance gaps identified only after production incidents occur. The enterprise AI agent estate has roughly doubled in four months, while security coverage has barely moved.

This is the exact pattern we saw in cloud computing circa 2014–2016: adoption sprinted ahead of governance, and the market response was a wave of standalone cloud security and compliance products — a dynamic that directly parallels why agentic automation programs fail when they skip the governance planning phase. Now we’re seeing the agent-era equivalent.

The trigger: agent sprawl

The term “agent sprawl” — agents being created faster than central teams can monitor or govern them — has entered the enterprise lexicon in 2026. Fewer than one in five organizations maintain a complete, current inventory of their AI systems, according to IBM’s “AI in Motion” research. Every major platform — Microsoft Copilot Studio, Salesforce Agentforce, AWS Bedrock, Google Vertex AI, UiPath Maestro — now makes it trivially easy for line-of-business teams to build agents. The Center of Excellence that took six months to stand up an RPA bot now faces a world where a product manager can deploy an agent in an afternoon.

The result: governance can no longer be a feature bolted onto the agent platform. It has to be standalone infrastructure that sits above the fragmented landscape of LLM frameworks and cloud providers, standardizing oversight from the outside in.

The Four-Layer Agent Governance Model

The four products that shipped between August and September 2026 map neatly to four distinct governance layers. Understanding these layers is the key to understanding why these products are complementary rather than competitive:

Each layer addresses a different failure mode. An agent with proper identity but no runtime policy can authenticate correctly and then do the wrong thing. An agent with runtime controls but no business KPI tracking can execute correctly while delivering zero value. A complete governance stack needs all four.

Layer 1: Identity — Okta Agent SSO

What it does

Okta Agent SSO, generally available since August 24, 2026, treats every AI agent as a first-class identity alongside human employees. It replaces static API keys — the single biggest identity anti-pattern in current agent deployments — with short-lived, identity-governed tokens using the open Cross App Access (XAA) standard.

Ric Smith, Okta’s president of products and technology, framed it directly: “AI agents are fast becoming a primary interface for how work gets done, but granting them access to enterprise systems shouldn’t require trading away security or visibility.”

How it works

Agent SSO registers an AI agent as a named identity in Okta’s Universal Directory, issues short-lived tokens scoped to specific resources and permissions, and enforces the same SSO policies (MFA, conditional access, session expiry) that govern human identities. When an agent’s token expires or its permissions change, it re-authenticates — exactly like a human session.

What makes it significant for RPA/automation teams

RPA developers have long dealt with the “service account” problem: bots running under shared credentials with overly broad permissions. Okta Agent SSO extends the identity-governance model that enterprises already use for human SSO into the agent world. If your Center of Excellence already manages Okta, this is a natural extension — no new vendor, no new procurement cycle. The product is included at no additional cost for existing Okta SSO customers.

The broader product: Okta for AI Agents

Agent SSO is part of the larger “Okta for AI Agents” suite, generally available since May 2026, which adds shadow agent discovery — finding unregistered agents running in the environment — and lifecycle management with named human owners assigned to each agent. This discovery capability directly addresses the agent sprawl problem.

Layer 2: Orchestration — IBM watsonx Orchestrate AgentOps Agent

What it does

IBM’s AgentOps Agent, GA since August 31, 2026, approaches governance from the orchestration layer. Rather than just monitoring agents, it actively improves them through an automated observe-test-evaluate cycle. Its three core components shipped simultaneously:

• AI Gateway: Discovers agents built on Amazon Bedrock and imports them into a single control plane (Azure AI Foundry and Google Vertex AI support arriving by end of September 2026)
• Trace Inspector: Follows an agent’s full execution path step by step — every tool call, every LLM invocation, every decision branch — giving teams forensic-level visibility into agent behavior
• Custom LLM-as-a-Judge Evaluation: Lets companies define what “good” means on their own terms using customizable evaluation criteria, rather than relying on generic quality scores

The AgentOps Agent itself

What sets IBM’s approach apart is that the governance tool is itself an agent. The AgentOps Agent handles the full improvement cycle through natural language conversation:

1. Observability discussion: You ask “How are my agents performing?” and it analyzes traces, latency, error rates, and output quality across your agent fleet
2. Test case creation: It builds test cases automatically from observed production patterns
3. Evaluation through user simulation: It runs your agent against many simulated conversations before changes reach production

This is a governance paradigm shift — from dashboards you stare at to an agent that proactively tells you what’s broken and can fix it.

Cross-platform reach

The AI Gateway’s ability to discover and import agents from external platforms (starting with AWS Bedrock) is critical. Most enterprises don’t build all their agents on one platform. The governance tool that can see across platforms wins — a key advantage IBM has been building toward with the watsonx Orchestrate Agentic Control Plane.

Layer 3: Security and Runtime — Broadcom AgentMinder

What it does

Broadcom unveiled AgentMinder at VMware Explore in Las Vegas on August 31, 2026, and it was generally available immediately. Where Okta answers “who is this agent?” and IBM answers “how is it performing?”, AgentMinder answers the security question: “What is this agent allowed to do right now, in this context, with this specific action?”

AgentMinder acts as what Broadcom calls a “traffic controller” for autonomous agents. It verifies each agent’s identity and intended mission, then authorizes every individual action against that mission before the action reaches enterprise resources.

Three core governance capabilities

1. Identity and Intent Controls: Treats autonomous agents as enterprise-grade identities and binds their authority to a declared mission, permitted intents, approved tools, and authorized resources. This goes beyond simple authentication — it’s mission-scoped authorization.
2. Runtime Policy Enforcement: A cloud-native gateway evaluates real-time context — identity, tool, intent, resource, current risk level — and enforces least-privileged policies on every tool invocation. An agent approved for “read customer data” can’t suddenly “write customer data” without a policy change.
3. Compliance-Grade Auditability: Built on OpenTelemetry, AgentMinder delivers chain-of-custody logging, anomaly detection, and operational insight at machine speed for every agent session and action.

Enterprise scale

Broadcom’s multi-region, active-active architecture handles peak loads of nearly 36 million customer-related and 7 million workforce-related API calls daily. This matters because runtime policy enforcement adds latency to every agent action — at this scale, the enforcement has to be measured in single-digit milliseconds or it becomes a bottleneck.

Why RPA architects should care

AgentMinder’s “mission-scoped authorization” concept is directly analogous to the process-specific permissions that well-governed RPA programs assign to each bot. If you’ve built a governance framework for RPA bots — where each bot has a defined process scope, approved applications, and audit trails — AgentMinder extends that same model to AI agents, with the addition of real-time intent verification that rule-based bots never needed (because they couldn’t deviate from their scripts).

Layer 4: Observability — Dataiku Agent Management

What it does

Dataiku Agent Management, announced September 24, 2026, with general availability planned for October, tackles the question the other three products don’t: “Is this agent actually delivering business results?”

This is a standalone product — it does not require the Dataiku platform — that provides a single control tower to track, evaluate, and govern agents across any platform. The key differentiator: it measures agents against business KPIs, not just technical metrics like uptime, latency, and error rates.

As Dataiku put it: “Technical monitoring tells you agents are running. It does not tell you they are doing their jobs.”

Cross-platform connectivity

Agent Management connects natively to agents running on:

• AWS Bedrock
• Databricks Agents
• Google Vertex AI
• Microsoft Copilot Studio and Azure AI Foundry
• Salesforce Agentforce
• Snowflake Cortex
• LangChain
• ServiceNow
• n8n
• Dataiku’s own platform

Plus OpenTelemetry support for custom environments. This is the broadest cross-platform coverage of any product in this analysis.

Core capabilities

• Agent inventory and discovery: Finds every agent across connected platforms, whether anyone registered it or not
• Business KPI tracking: Tracks performance against user-defined business metrics — revenue impact, case resolution time, customer satisfaction — not just “is it running?”
• Behavioral drift detection: Automatically identifies when agents miss performance targets or drift from expected behavior, catching issues before they create downstream risk
• Risk tiering: Classifies each agent by risk level — agents handling customers, sensitive data, or live transactions get flagged for mandatory certification and scheduled testing
• Governance workflow triggers: Based on risk thresholds and regulatory requirements, automatically triggers governance workflows (human review, re-certification, decommissioning)

The business case

For automation CoEs that already track RPA bot ROI through metrics like FTE savings and error reduction, Dataiku’s approach will feel familiar. It extends the same “prove business value” discipline to AI agents — something most organizations have struggled with because agent outcomes are harder to measure than deterministic bot outputs. If your CFO is asking “What are we getting for all this agent spending?”, this is the product that generates the answer.

Bonus Layer: StackGen’s Autonomous Operations Factory

StackGen’s Autonomous Operations Factory, launched September 15, 2026, deserves mention as a fifth entrant that blurs the line between governance and operations. Rather than governing agents built on other platforms, StackGen provides its own governed agents — four specialized “Aiden” agents covering InfraOps, DevOps, observability, and SRE — that share a common operational context (the “Aiden World Model”) and run under one set of policy checks and approval flows.

The platform adds policy enforcement, approvals, auditability, and cost tracking while executing through existing enterprise tools. Available now for AWS, Azure, Google Cloud, and Oracle Cloud users, with the broader platform and Autonomy Index in preview.

StackGen represents a different architectural bet: instead of governing any agent, build governed agents that replace ungoverned ones. For DevOps and SRE teams already drowning in ungoverned automation scripts, this approach has appeal — but it doesn’t solve the cross-platform governance problem the other four products address.

Head-to-Head Comparison Table

What Gartner Says: The Guardian Agent Category

Gartner formalized this space in February 2026 with its first-ever Market Guide for Guardian Agents — a new analyst category for AI agents whose primary purpose is to monitor, govern, and constrain other AI agents.

The Guardian Agent category, as Gartner defines it, covers three capability areas:

1. Visibility and traceability of agent behavior
2. Identity and access management for agents
3. Cross-platform policy enforcement

Gartner predicted in June 2025 that Guardian Agents will capture 10–15% of the agentic AI market by 2030. That forecast is now looking conservative given the velocity of product launches.

One statistic from the Market Guide is particularly sobering: “Through 2028, at least 80% of unauthorized AI agent transactions will be caused by internal violations of enterprise policies — information oversharing, unacceptable use, or misguided AI behavior — rather than from malicious attacks.” The threat isn’t hackers exploiting agents. It’s well-intentioned agents doing the wrong thing because no one told them the rules.

This directly parallels the RPA experience. The most common cause of RPA bot failures isn’t malicious interference — it’s process exceptions the bot wasn’t designed to handle. Agent governance extends that lesson: the agent that confidently summarizes confidential HR data into a public Slack channel isn’t malicious, it’s ungoverned.

What This Means for RPA Teams and Automation CoEs

If you run an automation Center of Excellence, the formation of the agent governance stack creates both an opportunity and a threat.

The opportunity

RPA CoEs already have the organizational muscle for governance: process documentation, bot access management, audit trails, ROI tracking, change management. These are exactly the capabilities the agent governance stack productizes. The CoE that positions itself as the enterprise agent governance authority — not just the RPA bot governance authority — becomes more relevant, not less, as agents proliferate.

Concretely, this means:

• Extend your bot inventory to an agent inventory. If you’re tracking 200 RPA bots in a spreadsheet, that same discipline needs to cover the 50 Copilot Studio agents and 30 Bedrock agents that your business units deployed last quarter. Dataiku Agent Management and Okta’s shadow agent discovery can automate this.
• Extend your bot access governance to agent identity governance. If every RPA bot has a named service account with scoped permissions, every AI agent needs the same — or better, short-lived tokens via Okta Agent SSO.
• Extend your bot ROI framework to agent ROI. The CoE that can answer “how much value are our agents delivering?” retains its budget. Dataiku’s business KPI tracking is the product implementation of what most CoEs are doing in spreadsheets.

The threat

The threat is governance sprawl: managing the governance tools themselves becoming as complex as managing the agents they’re meant to control. If you adopt Okta for identity, IBM for orchestration monitoring, Broadcom for runtime enforcement, and Dataiku for business tracking, you’ve just added four new vendor relationships and four new integration points to your stack — exactly the kind of tooling proliferation that derails agentic automation programs.

The practical recommendation: start with the layer that maps to your most acute pain point, not the layer that sounds most impressive.

Decision Framework: Which Governance Layers You Need First

Use this framework to prioritize which layer of the agent governance stack to adopt first, based on your organization’s current maturity:

The one non-negotiable: Before anything else, build an agent inventory. You cannot govern what you cannot see. Both Okta (via shadow agent discovery) and Dataiku (via cross-platform agent inventory) can automate this. If you’re starting from zero, this is step one.

After the inventory, layer in governance based on your regulatory and risk environment. Regulated industries (financial services, healthcare) should prioritize Layer 3 (runtime enforcement with audit trails). Organizations under CFO pressure to prove AI ROI should prioritize Layer 4 (business KPI tracking). Organizations with large, fragmented agent estates should prioritize Layer 1 (identity consolidation).

Frequently Asked Questions

What is the agent governance stack?

The agent governance stack is an emerging set of standalone enterprise products that govern AI agents across platforms. It typically includes four layers: identity (who the agent is), orchestration (how the agent performs), security and runtime enforcement (what the agent is allowed to do), and observability (whether the agent is delivering business results). Products from Okta, IBM, Broadcom, and Dataiku each address a different layer.

Do I need all four layers of the agent governance stack?

Not immediately. Start with the layer that maps to your most acute pain point — typically agent inventory and discovery if you’re early in your agent journey, or runtime enforcement if you’re in a regulated industry. A complete governance stack is the end state, not the starting point.

How does agent governance differ from RPA bot governance?

RPA bots follow deterministic scripts and their behavior is predictable. AI agents make autonomous decisions, can deviate from expected patterns, and interact with LLMs that introduce non-determinism. This means agent governance needs real-time intent verification (Broadcom AgentMinder), behavioral drift detection (Dataiku), and evaluation systems that assess quality rather than just execution status (IBM AgentOps). The governance principles are the same — identity, access control, audit trails, ROI tracking — but the enforcement mechanisms are fundamentally different.

What is Gartner’s Guardian Agent category?

Guardian Agents is a Gartner-defined category (formalized in February 2026) for AI agents whose primary purpose is to monitor, govern, and constrain other AI agents. The category covers visibility and traceability, identity and access management, and cross-platform policy enforcement. Gartner predicts Guardian Agents will capture 10–15% of the agentic AI market by 2030.

Can I use these governance products with UiPath, Automation Anywhere, or other RPA platforms?

Yes, with caveats. Okta Agent SSO works with any agent that uses Okta for authentication, including UiPath’s agentic capabilities. Broadcom AgentMinder governs any agent accessing enterprise resources through its gateway. Dataiku Agent Management connects to a wide range of platforms including Salesforce, AWS Bedrock, and Microsoft Copilot Studio, though direct UiPath and Automation Anywhere connectors haven’t been announced yet. IBM’s AgentOps Agent currently covers watsonx-native and Bedrock agents.

Key Takeaways

• A new infrastructure category is forming. Between August 24 and September 24, 2026, Okta, IBM, Broadcom, and Dataiku shipped standalone agent governance products — four products, four weeks, four different layers of the same stack.
• These products are complementary, not competitive. Each addresses a different governance layer: identity, orchestration, security/runtime, and observability. Enterprises will likely need multiple layers, purchased from different vendors.
• Agent governance is now a standalone market. Gartner formalized the “Guardian Agent” category in early 2026, predicting it will capture 10–15% of the agentic AI market by 2030. That forecast may be conservative given current launch velocity.
• RPA CoEs are uniquely positioned. The governance disciplines that automation CoEs already practice — inventory management, access governance, audit trails, ROI tracking — map directly to what the agent governance stack productizes. CoEs that extend their mandate to cover all agents (not just bots) become more relevant, not less.
• Start with agent inventory. You cannot govern what you cannot see. Both Okta and Dataiku offer automated agent discovery. Build the inventory first, then layer governance based on your regulatory environment and risk profile.
• Beware governance sprawl. Adopting all four governance layers from four different vendors creates its own complexity. Prioritize based on your biggest pain point and expand deliberately.

References

1. Okta. “Okta Brings First-Class Identity to AI Agents with Agent SSO.” Press Release, August 24, 2026. okta.com
2. IBM. “New in IBM watsonx Orchestrate: Cross-Platform Agent Discovery, Custom Evaluation and AgentOps Agent Goes GA.” August 31, 2026. ibm.com
3. Broadcom. “Broadcom Unveils AgentMinder, An Enterprise Solution for AI Agent Governance and Runtime Control.” Press Release, August 31, 2026. broadcom.com
4. Dataiku. “Dataiku Launches Agent Management for Enterprise AI Agents.” September 24, 2026. dataiku.com
5. StackGen. “StackGen Launches Autonomous Operations Factory to Govern Production Agents.” SiliconANGLE, September 15, 2026. siliconangle.com
6. Ellison, Dana. “The Agent Governance Stack Is Forming: Four Products, Two Weeks, One Pattern.” Yahoo Finance / Forkast News, September 12, 2026. yahoo.com
7. Gartner. “Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure.” May 26, 2026. gartner.com
8. Gartner. “Gartner Predicts That Guardian Agents Will Capture 10-15% of the Agentic AI Market by 2030.” June 11, 2025. gartner.com
9. Cloud Security Alliance. “CSA Research Note: AI Agent Governance Framework Gap.” April 3, 2026. cloudsecurityalliance.org
10. The Hacker News. “5 Learnings from the First-Ever Gartner Market Guide for Guardian Agents.” March 2026. thehackernews.com