Guardrails

Guardrails specify constraints on an agent’s behavior. Use them to identify actions it should avoid, conditions that should stop its work, or situations that require help from your team. Pass guardrails as a string in the agent’s configuration. The agent interprets these instructions while it works.

Set guardrails

This example keeps tCell’s configuration and adds constraints for testing your application:

import Antigen, { tCell } from "@antigen/sdk";

const antigen = new Antigen(process.env.ANTIGEN_API_KEY);
const original = await antigen.agents.get(tCell);

const agent = antigen.agent({
  ...original,
  guardrails: [
    original.guardrails,
    "Do not test endpoints under /internal/.",
    "Do not attempt denial of service.",
    "If you find exposed credentials, report them without using them.",
  ].join("\n"),
});

Combining the existing guardrails with your instructions preserves the constraints exposed in the original configuration. Passing a new string without the existing value replaces those configurable guardrails.

Write specific constraints

Describe the action or condition and what the agent should do in response.

Constraint	Example
Avoid an action	“Do not submit transactions through the production payment flow.”
Stop testing	“Stop the run when you confirm a critical vulnerability.”
Request human input	“If remediation requires a cloud configuration change, create a human task describing the change for the infrastructure owner.”
Limit proposed changes	“Do not modify database migration files when preparing a fix.”

Guardrails can apply to offensive testing, triage, or remediation. Choose instructions that match the work your agent performs. Use skills to explain procedures and provide context. Use guardrails to constrain how the agent carries out that work.

Update saved guardrails

Use antigen.agents.update() to change a registered agent’s guardrails. The string you provide becomes its saved value.

Add constraints

Retrieve the agent and combine its existing guardrails with the new instructions:

const saved = await antigen.agents.get("our-security-agent");

await antigen.agents.update("our-security-agent", {
  guardrails: [
    saved.guardrails,
    "Do not test endpoints under /internal/.",
  ].join("\n"),
});

Replace the guardrails

Pass a new string containing the configurable guardrails you want the agent to use:

await antigen.agents.update("our-security-agent", {
  guardrails: `
Do not test endpoints under /internal/.
Do not attempt denial of service.
If you find exposed credentials, report them without using them.
  `.trim(),
});

Replacing this field changes the configurable instructions. It does not remove a pre-built agent’s internal restrictions or the platform’s enforced controls.

Guardrails and enforced restrictions

Guardrails guide the agent through natural-language instructions. They are not a substitute for access controls or isolation. Target approval and sandbox restrictions are enforced separately. A guardrail cannot authorize testing an unapproved target or grant access the agent does not have.

Getting started

Asset Map

Runs

Vulnerability Lifecycle

Team Workflows

SDK

CLI

Integrations

Set guardrails

Write specific constraints

Update saved guardrails

Add constraints

Replace the guardrails

Guardrails and enforced restrictions

​Set guardrails

​Write specific constraints

​Update saved guardrails

​Add constraints

​Replace the guardrails

​Guardrails and enforced restrictions

Set guardrails

Write specific constraints

Update saved guardrails

Add constraints

Replace the guardrails

Guardrails and enforced restrictions