Isn't a human reviewer enough?

It is, until volume grows. Once an agent produces hundreds of outputs a day, full human review stops being viable and becomes a rubber stamp. The supervisor filters the obvious 90% so the person only sees what genuinely needs judgement.

Can the same model supervise itself?

Poorly. If it shares prompt, context and data with the supervised agent, it tends to validate its own errors. Useful supervision uses explicit criteria, different sources and, where risk justifies it, a different model.

When is a human still required?

Three cases: when the decision commits money above a threshold, when accountability cannot legally be delegated, and when the case resembles nothing the policy anticipated.

← Full glossary

Supervisory agent

governance

guardian agent · verifier agent · automated second line of defence

An AI agent whose only job is to review what another agent does before the action ships: it checks the output against company policy, flags what is missing, and decides whether it passes, is blocked, or escalates to a person.

A supervisory agent is an automated second line of control. It does not do the work; it reviews it. It runs alongside the producing agent (or between that agent and the target system) and evaluates every output or action against three things: written company policy, approved sources of information, and the authority limits set for the supervised agent. The outcome is always one of three decisions — allow, block, or escalate to a human — plus a record of why. Its value is not catching quality defects (offline evaluations do that) but stopping a wrong action from reaching a customer, a regulator, or a system of record. In regulated sectors it is used to verify mandatory disclosures, conflicts with internal policy, and claims that no approved source supports. Design requirement: a supervisory agent must not share context or prompt with the agent it supervises, or it will repeat its mistakes; and it does not replace the human on decisions where accountability cannot legally be delegated.

How it differs from

Human in the loop: Human in the loop is a person approving. A supervisory agent is software filtering before it reaches that person. They combine; they do not replace each other.
Agent guardrails: Guardrails are fixed rules in the agent's configuration. The supervisor is a separate agent that reasons over the output with context and can escalate.
Agent observability: Observability tells you what happened afterwards. The supervisor decides before it happens.
Agent evals: Evals measure quality against a test set, outside production. The supervisor acts in production on live traffic.

Examples

An agent reviews every sales email another agent drafts and blocks those missing the mandatory legal disclosure.
Before a support agent promises a refund, a supervisor checks the amount against the returns policy and escalates if it exceeds the limit.
A supervisor cross-checks the figures a marketing agent cites against the single approved data source and returns the piece if they do not match.

FAQ

Isn't a human reviewer enough?: It is, until volume grows. Once an agent produces hundreds of outputs a day, full human review stops being viable and becomes a rubber stamp. The supervisor filters the obvious 90% so the person only sees what genuinely needs judgement.
Can the same model supervise itself?: Poorly. If it shares prompt, context and data with the supervised agent, it tends to validate its own errors. Useful supervision uses explicit criteria, different sources and, where risk justifies it, a different model.
When is a human still required?: Three cases: when the decision commits money above a threshold, when accountability cannot legally be delegated, and when the case resembles nothing the policy anticipated.

Gartner Market Guide for Guardian Agents (via vendor distribution) · 2026-09-07 · E3

Supervisory agent

How it differs from

Examples

FAQ

Related terms