Worknano

AI agent guardrails, with examples

Updated 2026-11-15

AI agent guardrails are fixed limits around an AI agent that keep a mistake small. An AI agent is a program that uses AI to carry out tasks for you. Guardrails decide what the agent gets to see, what it may do and when it must hand over to a person. Good guardrails do not rely on the AI behaving well. They hold even when it gets something wrong.

An instruction is not a guardrail

Think of a company card. You can tell an employee: "Please do not spend more than 500 euros." That is an instruction. Or the bank can set a limit of 500 euros on the card. That is a guardrail. The first depends on the person. The second works whatever happens.

It is the same with an agent. "Never offer more than ten per cent discount" in its instructions is a request. The agent will follow it nearly always. A guardrail is the check that stops a quote with twelve per cent, whatever the agent wrote. Write the instruction as well, but do not count it as protection.

Four kinds of guardrail, with examples

A guardrail set for one agent

Here is what a full set can look like for an agent that drafts replies to customer requests.

  1. It only receives requests that came in through the contact form.
  2. It reads that customer's orders and earlier requests, and your help texts. Nothing else.
  3. It writes a draft. It cannot send.
  4. The draft must name the help text or the order it relies on.
  5. It never promises a date, a refund or a price change. Where one is needed, it leaves a marked gap for the person.
  6. If it cannot answer from what it read, it says so and passes the request on.

Six rules, and each one can be checked. None of them asks the agent to "be careful".

How to write a guardrail that works

How many guardrails are enough?

Start strict and loosen step by step as the agent proves itself. With too few guardrails, one bad run does damage. With too many, the agent hands everything over, and your team has the old workload plus a queue. The balance is right when hand-overs are a small share of the work and most of them turn out to be justified.

How Worknano handles this

In Worknano, what an agent can reach comes from two places: the tools you give that agent, and the role of the person who asks. Changes to records appear as a preview to confirm. For limits that depend on an amount, approval chains can include a step such as a finance approval only above an amount you choose. Workflow conditions decide which records reach an agent at all.

Create a free workspace and set the first limits for an agent of your own.