AI agents · Types
What is a guardian agent?
The short answer
Its job is to watch, check or stop other agents. Gartner names three roles: reviewers that check output, monitors that track actions, and protectors that adjust or block an action as it happens.
It is one of the 4 forms of multi-agent system, with the agent pipeline and the manager agent and the agent team.
| What starts it | Another agent's work |
|---|---|
| Who decides the next step | The guardian, against rules you set |
When should you use a guardian agent?
Use it when agents act on customers, money or regulated records, or when more agents are running than people can review.
When should you not?
Not when it shares the instructions and the blind spots of the agent it checks. It then adds cost and no independence.
What does it need to run safely?
It needs independence from what it checks, the authority to block and not only to report, and a record of its own.
An example
Before a customer letter drafted by another agent is sent, the guardian checks it against the disclosure rules and holds the ones that fail.
What else is it called?
Reviewer, critic, evaluator, monitor. Google ships a sample agent that monitors other agents in production. Gartner expects guardian agents to take 10 to 15 percent of the agentic AI market by 2030.
How do you build one on Katonic?
- Control Room, Guardrails. Choose Create Profile. A profile holds three kinds of rail: Input rails "applied before the prompt reaches the LLM", Output rails "applied to the LLM response before it leaves the system", and Tool rails "applied to MCP tool-call arguments". Each rail acts by Redact, Block or Warn.
- In each agent's Guardrails section, pick that profile. The runtime reads the profile, not the fields on the agent.
- Control Room, Policies. Choose Create Policy to require approval for a tool call, or to block it, by tool and scope. Policies run in priority order.
- Control Room, Approvals. Held tool calls wait here for a named approver, with an audit trail of who decided.
- In an agent's Guardrails section, use Test against a prompt to see which guardrails would trip before you deploy.
Questions people ask
What starts a guardian agent?
Another agent's work.
Who decides its next step?
The guardian, against rules you set.
What is a multi-agent system?
A multi-agent system is several agents, each with its own instructions and tools, working on one piece of work. The forms differ in who decides the order. OpenAI advises starting with one agent and moving to several only when needed. Microsoft's guidance is to start with several where the work crosses security boundaries or teams, and otherwise to test a single agent first. Gartner expects 70 percent of AI applications to use multi-agent systems by 2028.