Guardian agents are technologies that supervise AI agents: they observe what agents do, check it against policy, and adjust or block actions that go out of bounds. The term comes from Gartner, which treats guardian agents as the oversight layer for agentic AI. In practice, a guardian is any independent control point that sits between an agent and the systems it acts on, which is why it is often called the AI agent governance layer.
How does Gartner define guardian agents?
Gartner introduced the idea publicly in a June 2025 press release. It described guardian agents as technology that helps keep interactions with AI trustworthy and secure, and named three kinds of use:
- Reviewers check AI-generated output and content for accuracy and acceptable use.
- Monitors observe and track AI and agent actions so that people or other systems can follow up.
- Protectors adjust or block agent actions and permissions automatically while they run.
The same release forecast that guardian agent technologies would make up at least 10–15% of the agentic AI market by 2030.
In February 2026 Gartner published its first Market Guide for Guardian Agents, led by analyst Avivah Litan. The full guide is for Gartner clients, so what follows comes from public coverage, not the guide itself. IT Brew describes the category as a mix of AI governance and runtime controls that secure agent activity. It reports Gartner's estimate that companies spend under 1% of their agentic AI budgets on guardian agents today, rising to 5–7% by 2028. A partner-contributed summary in The Hacker News, written by a vendor, lists three mandatory capability areas: AI visibility and traceability, continuous assurance and evaluation, and runtime inspection and enforcement. It says Gartner sees most deployments as prototypes or pilots, and calls for an independent guardian layer that works across clouds and platforms rather than inside one vendor's stack.
Treat these secondhand details as provisional until you can read the guide.
How do guardian agents relate to AI TRiSM?
AI TRiSM is Gartner's term for AI trust, risk and security management: the framework it uses for the governance, trustworthiness, reliability, security and data protection of AI systems (Gartner glossary). Gartner's public descriptions split it into layers, from AI governance at the top down to runtime inspection and enforcement, information governance and the underlying infrastructure. Guardian agents are the agentic version of the runtime layer: where AI TRiSM asks for controls that inspect and enforce while AI is running, a guardian applies them to what agents do. A policy-enforcing gateway supplies runtime enforcement and an audit record that feeds the governance layer; it does not replace model evaluation, data governance or an AI inventory.
How are guardian agents delivered?
Per the same summary, Gartner lists six delivery approaches: standalone oversight platforms, AI and MCP gateways, embedded or in-line runtime modules, orchestration layer extensions, hybrid edge-cloud models, and coordination mechanisms such as standards and APIs. A gateway only works if agent traffic actually has to pass through it, so network design matters as much as the product.
| Observe-only (monitor) | In-line enforcement (protector) | |
|---|---|---|
| When it acts | After the action, from logs and traces | Before the action reaches the tool |
| Can it stop a bad call? | No; it alerts someone | Yes: deny, redact or hold for approval |
| Typical form | Posture tool, SIEM rules, dashboards | Gateway or in-process runtime module |
| Main risk | Too late for destructive actions | Agents that can route around it |
Does a guardian agent have to be an LLM?
No. The word "agent" suggests a model watching a model, and some products do work that way: a second model reviews the first model's output or plans. That helps with judgement calls, such as whether an answer is accurate or on-topic. But a model-based checker can itself be wrong or manipulated, and its decisions are hard to reproduce or explain to an auditor.
For actions, deterministic policy is usually the better guardian. "This agent may not call delete_* on production without two approvers" is a rule, not a judgement. It gives the same answer every time, can be tested in CI, and leaves a record that names the rule. Many teams will want both: deterministic enforcement on actions, and model-based review where judgement is genuinely needed.
Where does MAQPNA fit as a guardian layer?
MAQPNA is a self-hosted runtime for AI agents whose gateway acts as an in-line, policy-enforcing guardian. It is not itself an LLM. Every MCP tool call, agent-to-agent call, model call and egress connection goes through the gateway, and each agent session runs in a sandbox whose only network exit is the gateway. Mapped loosely onto Gartner's three capability areas:
| Capability area | What MAQPNA does | Docs |
|---|---|---|
| Runtime inspection and enforcement | Default-deny policy per tool and argument; allow, deny or hold for human approval; in-process DLP; taint rules after untrusted input; budgets; a kill switch | Policies, approvals, kill switch |
| Visibility and traceability | A signed identity per session naming the agent and the person it acts for; a hash-chained audit ledger of every decision; OpenTelemetry traces | Audit ledger, identity |
| Continuous assurance and evaluation | Policies tested offline in CI, replayed against past traffic and rolled out in dry-run mode first; posture checks with maqpna doctor | Write policies |
Decisions come from the policy engine: native rules, with optional Cedar or Rego that can only make a decision stricter. DLP uses deterministic detectors and sends nothing to an outside classifier. You can add optional prompt-injection classifiers at the gateway; they fail open by default, and the docs say so. Because MAQPNA runs on your own cluster with no vendor control plane, it is independent of any single model or cloud provider, which is the kind of neutral layer Gartner's framing asks for.
What does MAQPNA not cover?
Be clear about the edges. MAQPNA does not judge whether a model's answer is accurate, so it is not a "reviewer" in Gartner's sense. It does not include behavioural anomaly scoring; you would run that in your SIEM on the ledger stream. And it cannot stop a model being fooled; it limits and records what a fooled agent can do. The security model lists these limits.
Related: AI agent governance in practice, what is an MCP gateway, the AI agent kill switch, OWASP Top 10 for Agentic Applications mapping and sovereign, air-gapped agent governance.