Supervisory AI is the category of AI systems whose function is to oversee other AI systems – discovering them, observing their behavior, evaluating their actions against policy, and intervening when intent diverges from authorized scope. It is the umbrella unifying discovery, observation, and intervention as the three core functions of an AI control plane.
- AI that oversees other AI rather than performing the primary task
- Unifies discovery, observation, and intervention into one control function
- Reviews agent decisions at machine scale, escalating only what needs a human
- Gartner describes the practitioner category as guardian agents
Why is supervisory AI important?
Supervisory AI matters because the agent-to-human ratio breaks every program that relies on people reviewing agent decisions. A security and governance team cannot personally examine what a few thousand agents do across a working week, and sampling gives you confidence proportional to the sample rather than the population.
The supervisory layer is the system that does the reviewing at machine scale. It watches every session, evaluates each action against policy, surfaces the decisions a human reviewer should see, and intervenes on the ones policy will not permit. Human judgment stays in the loop where judgment is what is needed, rather than being spent on volume.
The alternative is one of two failures. Either the program reviews a fraction and reports on that fraction, or it slows agent adoption to whatever pace human review can sustain. Gartner has begun describing this practitioner category as guardian agents, a signal that the oversight layer is being treated as its own discipline rather than a feature.
What is supervisory AI?
Supervisory AI is any AI system whose purpose is to govern other AI systems rather than to perform a primary business task. Its subject is agent behavior, and its output is a decision about that behavior.
Four functions define it. Discovery establishes which AI systems exist across the environment. Observation captures what those systems do – prompts, reasoning steps, tool calls, and data access – as it happens. Evaluation compares each action against policy and against an established behavioral baseline. Intervention acts before an action completes, allowing, redirecting, modifying, or stopping it.
The reason these belong together is that each depends on the one before it. Observation without discovery watches a subset. Evaluation without observation judges summaries. Intervention without evaluation is a blunt block list. A supervisory system holds all four on the same data model, which is what lets it act on what it sees rather than report on it.
Supervisory AI is not a model guardrail. It operates on the agent and its actions rather than on the text a model produces.
Types of supervisory AI
Supervisory systems differ by how much authority they hold and by what they supervise.
By authority, observational supervisors watch and report, which suits early deployments and produces no enforcement. Advisory supervisors recommend an action and wait for a human decision, useful where consequence is high and volume is low. Autonomous supervisors act within a defined policy envelope and escalate only what falls outside it, which is the only model that scales to large agent populations.
By subject, session supervisors evaluate a single agent's activity in context. Fleet supervisors evaluate patterns across many agents, catching coordinated or emergent behavior no single session reveals. Escalation supervisors handle the ambiguous cases the primary layer flags, applying deeper analysis to a small number of hard decisions rather than fast evaluation to all of them.
Mature deployments run more than one of these, with the escalation layer reserved for what the fast path cannot resolve.
Supervisory AI & Onyx
Onyx is built as a supervisory AI platform. The Secure AI Control Plane discovers every agent, observes every action, evaluates every request against policy, and intervenes at the request boundary. The Guardian Agent escalation layer inside AI Guard handles deep analysis for the ambiguous or complex threats the fast path flags rather than resolves.
Steering shapes how that oversight lands. An action can be redirected or modified at action time rather than only blocked, so a task completes within policy instead of failing at the boundary. Just-in-Time access narrows what any action could reach, evaluating context at the request boundary rather than at role assignment. Underneath both, AI-native controls were built for the agent layer from the start rather than adapted from human-user tooling. Together they are what allow oversight to scale without turning every marginal action into failed work.
