Skip to content

Guardian agents: what runtime oversight of AI agents actually requires

13 Nov 2026

·

6 minute read

Guardian agents are a supervisory layer that watches another AI agent while it runs and can intervene before an action lands. The idea has arrived quickly and the term is used loosely, so it is worth separating what the research proposes from what the word is being attached to. The short version: a guardian watches live behaviour, because the failures worth catching do not exist until the agent is running.

The failures that only exist at runtime

The argument for watching an agent during execution is that the serious problems cannot be found any earlier.

The MI9 framework, a runtime governance proposal for agentic systems, makes this its starting point: the most critical risks "arise dynamically and elude pre-deployment control methods", and static pre-deployment assessment cannot capture behaviour that only emerges during execution. Its authors name four in scope:

  • Goal drift under a fixed stated goal, where the objective never changes on paper but the agent's working interpretation of it does.
  • Policy-skipping tool chains, where each individual call is permitted and the sequence achieves something that is not.
  • Delegated privilege escalation, where an agent hands work to another component that holds more access than it does.
  • Multi-agent coordination failures.

Read those again and notice what they have in common. None is visible in a single tool call. All of them are properties of a sequence, which is why a test that checks one action at a time will pass while the behaviour is wrong. MI9 describes safety properties enforced "over live behavior sequences", and that phrase is the whole design constraint.

What a watching layer has to track

MI9 sets out six mechanisms, and they are a reasonable checklist for anyone building or buying this:

  1. An agency-risk index that scores autonomy, adaptability and continuity, so oversight is proportionate rather than uniform.
  2. A telemetry schema that captures cognitive, action and coordination events, because an action log alone cannot show drift.
  3. Continuous authorisation monitoring that adjusts permissions to the agent's current context.
  4. A conformance engine using finite-state machines to enforce multi-step policy patterns.
  5. Behavioural drift detection against goal-conditioned baselines.
  6. Graduated containment, escalating interventions that preserve operational continuity instead of halting everything.

The second and sixth are the ones teams underestimate. Telemetry that records only tool calls cannot detect goal drift, because drift happens in the reasoning between calls. And a guardian whose only move is to stop the system will be switched off within a month, because an oversight layer that takes the service down is more expensive than the risk it prevents.

One watcher is not enough once agents talk to each other

A single supervising agent has a specific blind spot, and the research is direct about it.

QuadSentinel, which compiles natural-language policies into machine-checkable rules enforced at runtime, describes itself as "replacing a single supervisor with a multi-agent guard team", splitting the work across a state tracker, threat watcher, policy verifier and referee, with the referee issuing allow or deny decisions and returning its reasoning to the agent so it can correct itself.

Its criticism of the alternative is worth quoting, because it explains why many existing guardrails miss things. Those guardrails "treat agents as isolated black boxes", and a tool-triggered guard such as ShieldAgent fires only on tool invocations, "making it blind to the inter-agent communication layer". The paper's conclusion is that "single-agent guardrails fall short" in decentralised multi-agent systems.

If your system is one agent calling tools, a tool-level guard covers most of the surface. Once agents message each other, the messages are where the coordination failures live, and a guard that never reads them is watching the wrong channel.

A guardian detects, the permission boundary enforces

This is the part to be careful about, because it is where a guardian agent can give false comfort.

A guardian is itself a model-driven component making judgements, and OWASP's guidance on excessive agency is explicit about where enforcement belongs: "Implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not." The same page puts the root cause of excessive agency as "typically one or more of: excessive functionality; excessive permissions; excessive autonomy", none of which a watcher fixes.

So the division of labour is this. The refund tool rejects a call outside policy whatever any agent concluded. The guardian catches the patterns a single permission check cannot see, like a sequence of individually legal calls that adds up to something nobody authorised. Adding the second while skipping the first produces a system that is observed and still unsafe. We have written about the boundary itself in what a production agent needs before it is safe to use in a business.

Most teams need the cheaper things first

The honest position on guardian agents today is that the research is young, the products are younger, and a supervisory layer is not the first control worth building.

In order:

  1. Least privilege on every tool, so the worst outcome of a wrong decision is bounded.
  2. Authorisation enforced in the downstream system rather than in the model's judgement.
  3. An evaluation set built from real failures, so you know what the agent gets wrong before you automate watching for it. Building one from your own failure cases takes a week.
  4. Telemetry rich enough to reconstruct a run, which is a prerequisite for any guardian you later add, and useful on its own.
  5. A guardian layer, once you have multiple agents coordinating and the sequence-level failures above are a real possibility rather than a slide.

Teams that do the first four rarely regret it, and they arrive at the fifth knowing what their guardian should watch for, which is the only way to configure one sensibly.

If you are designing oversight for agents that touch money, patient records or student records, tell us about the project.

AI tools

Ask ChatGPTAsk Claude

Start with the problem that matters