Skip to main content

LiveStart free with 200 credits

Product

Agento Guardrails: Scope Checks, a Kill Switch and a Human Release for AI Agents

Agento Guardrails keeps an AI agent inside its job. Every step is checked against the agent's scope, anything outside it is refused, an admin can stop a run with one click and every refusal is kept on record.

AUArsalan UsmaniFounder5 min read
Agento Guardrails: Scope Checks, a Kill Switch and a Human Release for AI Agents

For security, IT and risk teams who need to know what an AI agent can touch, and how to stop it.

Agento Guardrails is the part of Agento that checks every step an AI agent takes against the agent's scope before the step runs. It refuses anything outside that scope, lets an admin stop a run and hold the agent, and records every refusal.

Key takeaways

  • Every step an agent takes is checked against its scope before it runs. The scope is the apps, files and actions the job needs.
  • A step outside the scope is refused before anything is read or sent. In the hiring run shown here, the agent reached outside its scope nine times and was refused nine times.
  • An admin can isolate a run with one click. The run is stopped for good, and the agent behind it can be put on hold.
  • A held agent only runs again when a person releases it.
  • Every refusal is counted and listed with the rule, the step, the action tried and the run it came from.

The problem: an agent that can reach further than its job

A hiring agent needs to read the résumés for one role. It does not need the rest of the document library, and it does not need to send email. Before an agent goes anywhere near production, a security team wants four answers:

  • What can it touch? Which apps, which files and which actions.
  • What happens when it tries something else? Whether the attempt is stopped, or only noticed afterwards.
  • How do we stop it, and who can start it again? A kill switch is only useful if the restart is controlled too.
  • Can we prove what happened? An auditor will ask for the record, not a description.

Guardrails answers each one. We have written before about why the blast radius of an agent matters. This page shows how the limit works in practice.

How it works: every step is checked against the agent's scope

1. Every run starts for a reason

A run begins in one of four ways: someone clicks Run, a schedule fires, a webhook arrives or a person asks in chat. Nothing starts on its own.

2. The agent has a scope

A scope has three parts. This is the scope of the hiring agent in the video, whose job is to screen the candidates:

Part of the scopeHiring agent
AppsMail and Drive
FilesRésumés for this role
ActionsRead, summarise, draft

3. Each step is checked before it runs

When a step is inside the scope, it runs. When it is outside the scope, it is refused. The check happens first, so a refused step never touches the app it was reaching for.

What a refusal looks like

Nine reaches outside scope, nine refusals

In the hiring run, the first step listed sites, listed files and searched files in SharePoint. It then tried to download files that were not among the step's allowed actions. Each download was blocked by a rule: nine reaches outside its scope, nine refusals. The run inspector shows every connector call the step made and marks each blocked one.

Blocked before anything was sent

In a second workflow, one step listed files in Google Drive and completed normally. The next step tried to send with Gmail. That send was refused, so nothing was sent.

Stopping a run: one click, and the agent is held

Each run has an evidence booklet with a Security controls section. From there an admin chooses Isolate run, gives a reason, and can add an incident reference. Three things follow:

  • The run is stopped for good. It is blocked from touching any connected app and cannot be resumed afterwards.
  • The agent can be put on hold. A hold pauses that agent in other runs of the workflow, along with the steps that depend on it. Branches that do not need the agent keep going.
  • The stop is on the record. The isolation, the reason, the incident reference, who did it and when are all kept with the run.

Starting again: only a person can release it

The next run of the workflow waits at the held agent. Its steps show as on hold, and they stay that way until a person releases the hold. The agent does not come back by itself.

Oversight: every refusal counted

The Guardrails console summarises the last 24 hours: runs finished, workflows that left their usual path, steps blocked by a rule and freeze decisions. In the video it shows 248 runs finished across 35 workflows, no workflow off its usual path, 10 steps blocked by a rule and nothing frozen.

Below the summary, each blocked step is listed with:

  • The rule that blocked it, for example a file download.
  • The step that made the attempt.
  • The action tried, for example reading a file in SharePoint.
  • The reason, for example "not among the step's allowed actions".
  • The run it came from.

That list is the answer to the auditor's question. Related controls are described on our security page and in the Trust Center.

Frequently Asked Questions

What is an AI agent guardrail?
A guardrail is a rule that is checked before an agent's step runs. In Agento the guardrail is the agent's scope: the apps it may use, the files it may open and the actions it may take.
How do I limit the blast radius of an AI agent?
Give the agent a scope that covers only what its job needs, and check every step against it before the step runs. In Agento a step outside the scope is refused, so a mistake or a manipulated prompt cannot reach apps or files the agent was never given.
Is there a kill switch for an AI agent in production?
Yes. An admin can isolate a run from its evidence booklet. The run is stopped for good and blocked from touching any connected app, and the admin can place a hold on the agent at the same time.
Who can restart an agent that has been held?
Only a person. The next run waits at the held agent until someone releases the hold.
What is recorded when a step is refused?
The rule, the step, the action that was tried, the reason and the run it came from. Isolating a run also records the reason given, any incident reference, who did it and when.
Can we see it on our own workflows?
Yes. Contact us to book a walkthrough, or sign up to start building workflows in Agento.
Back to all articles