Forrester predicts 25% of planned enterprise AI spending in 2026 will be deferred to 2027, specifically because CFOs are demanding ROI evidence and security teams are flagging governance gaps that haven't been closed. That's a hard, budget-level number, not a vague concern. And it's the clearest sign yet that governing AI agents has stopped being a feature vendors bolt on and started becoming a line item enterprises fund on its merits, with venture money and enterprise spending now flowing toward closing exactly the gap that's stalling deployment everywhere else.
Two distinct models are emerging to build that category, and they work in fundamentally different ways. One is a traditional control plane, rules, identity, and audit logs applied to a new kind of software. The other is stranger and more advanced: agents watching agents. It's happening in production already, and it comes with a problem worth knowing about before anyone assumes it's solved.
Why this became a category at all
Gartner's March 2026 report formally named the emerging space: the Agent Management Platform, described as a control plane sitting above individual agent frameworks, providing governance, observability, and cost intelligence that the frameworks themselves don't. OWASP gave agentic applications a dedicated Top 10 threat list this year, naming goal hijacking, tool misuse, and identity privilege abuse as the categories that matter most for autonomous systems. NIST launched a dedicated AI Agent Standards Initiative in February 2026, with identity management for AI agents named as one of six priority areas. When an analyst firm names the category, a security standards body writes a dedicated threat list for it, and a federal standards body opens a formal initiative on it, all within months of each other, that's a converging signal, not one vendor's marketing framing.
Model one: the control plane
The more traditional model treats agent governance the way IT has always governed software: identity, permissions, and an audit trail, applied to a new kind of process. Kore.ai launched what's described as the first purpose-built Agent Management Platform addressing cross-framework governance in March 2026, built to work across LangGraph, CrewAI, AutoGen, Google ADK, AWS AgentCore, Microsoft Foundry, and Salesforce Agentforce rather than locking a customer into one vendor's ecosystem. Google's Gemini Enterprise Agent Platform, launched at Google Cloud Next in April, is explicitly built to win the governance fight rather than the benchmark fight, placing its controls at the infrastructure layer instead of the application layer the way Anthropic and OpenAI have mostly done. IBM's watsonx Orchestrate and UiPath's Maestro both describe themselves in similar terms: a control plane that deploys, audits, and enforces policy across potentially thousands of agents from any source.
The pattern underneath all of these is a four-part framework showing up across the industry: agent identity, runtime enforcement, comprehensive auditing, and lineage provenance, tracking exactly which data an agent's output traced back to. None of this is exotic. It's the same governance logic that's existed for human employee access for years, applied to software that now acts with a comparable degree of independence.
Model two: standing guard, agents keeping agents in line
The second model is a different thing entirely. Rather than a rules-based control plane sitting outside the agents, an AI system actively evaluates another AI agent's behavior as it happens. The pattern is called LLM-as-a-Judge in its simpler form, and Agent-as-a-Judge in its fuller one. One enterprise example: a customer support platform running multiple AI agents in production, an AI Support Agent, an AI Knowledge Agent, a Case Quality Auditor, all orchestrated together, with an LLM judge supervising, comparing, and ranking their outputs for accuracy and compliance, every judgment logged with a matching audit trail. The academic version goes further still: Rather than a single model scoring a final answer, a full reasoning agent evaluates another agent's entire multi-step process, the intermediate decisions and actions along the way, not just what it ultimately produced.
This is a more capable form of oversight than a fixed rule can offer, and it's already running in production today, not a research curiosity.
That finding is the reason oversight researchers increasingly argue that monitoring has to be designed as an external constraint, not just an internal instruction the agent is trusted to follow in good faith. An agent told to behave a certain way, and an agent whose behavior is checked against a boundary it cannot talk its way around, are not the same guarantee.
Where this lands
Both models are useful, and both are worth watching. But the evasion finding is a specific point in favor of the pattern running through every guide in this Knowledge Library: AI drafts, a deterministic rule checks it, a human sends, everything logs. A fixed rule and a human reviewer aren't something an agent can learn to talk its way past the way it can learn to satisfy another AI's judgment. The control-plane model and the agent-judging-agent model are both meaningful advances in how AI gets governed. Neither replaces the value of a boundary that doesn't reason, and a person who still decides.
Sources: AI Pulse · Big Picture · workplaceai.ai. Forrester's 25%-deferred prediction and Google's Gemini Enterprise governance positioning: TechTimes, July 19, 2026. Gartner's Agent Management Platform category and Kore.ai's March 17, 2026 launch: Kore.ai's blog, citing Gartner's March 2026 report. OWASP's Top 10 for Agentic Applications 2026 and the four-part governance framework: Promethium's AI Agent Data Governance enterprise playbook, April 24, 2026. NIST's AI Agent Standards Initiative: Federal Register request for information, January 8, 2026. IBM watsonx Orchestrate and UiPath Maestro: Agentic AI Knowledge Base's enterprise platforms overview. LLM-as-a-Judge in production: SearchUnify, May 6, 2026. Agent-as-a-Judge academic framework: "A Comprehensive Survey of Self-Evolving AI Agents," arXiv, citing Zhuge et al. 2024. The agent-evasion finding: "AI Agents Under EU Law: A Compliance Architecture for AI Providers," arXiv, April 6, 2026, citing Kim et al.'s survey of adversarial research. EU AI Act audit-log retention requirement: as reported by IndextDataLab, April 24, 2026. Every figure above is attributed to its original researcher; none is a WorkplaceAI study.