- Getting started
- Prerequisites
- Building agents in Studio Web
- Building agents in Agent Builder
Agents user guide
Guardrails
Guardrails provide a mechanism to control unexpected behaviors within tool calls deterministically. They allow developers to configure conditions for human intervention and ensure that escalations occur precisely when defined conditions are met.
Briefly, the role of guardrails is to:
- Address unpredictable tool call inputs and outputs at runtime.
- Reduce the need for human-in-the-loop (HITL) interventions for correction and validation tasks.
- Enable developers to create custom rules at the individual agent/tool level.
You define guardrails for each agent tool, as follows:
-
There can be one or more guardrails per tool.
-
Each guardrail contains one or more rules and one action. Rules are evaluated against the inputs and outputs of the tool.
-
An action is triggered when all rules are met. The action can happen before and/or after tool execution.
Guardrails are displayed in the Properties panel for every type of tool (processes, activities, or other agents). The Guardrail builder is where you define the list of rules that trigger the guardrail and the action to be applied when all the rules are met.
At runtime, guardrails are checked from top to bottom in the order they appear in the list. You can reorder guardrails as necessary by a drag and drop action using the menu icon on the left.
Use the Enable guardrail for evaluations option to apply the selected guardrail during evaluation runs as well.
To configure guardrails for a tool:
- Add any tool to your agent.
- Select the tool and open the Properties panel.
- Select Open guardrail builder.
- Set a guardrail name and description, then configure the rule and action types.
- Select Save. Tools with guardrails are marked with a shield icon.
Rule types
The following rule types are available:
- Always enforce the guardrail (default) – This guardrail always triggers the action regardless of input or output. It is applicable to the agent input (pre-execution), output (post-execution), or both (pre and post-execution).
- String – This guardrail applies to available input and output fields. The rule is applicable to string fields, including nested JSON fields, and supports various operators (contains, starts with, ends with, equals, is empty, etc.)
- Number – This guardrail applies to available input and output fields. The rule is applicable to number fields, including nested JSON fields, and supports various operators (contains, starts with, ends with, equals, is empty, etc.).
- Boolean – This guardrail applies to available input and output fields. The rule is applicable to boolean fields, including nested JSON fields, and supports the Equals operator, with True or False values.
Action types
For each rule, you enforce the action to occur when all the rules are met.
The following action types are available:
- Log – Sets a severity level for logging. This action is useful for sending custom events for dashboard monitoring
- Filter – Removes selected
fields from the input and/or output of a tool, as follows:
- Inputs: Selected fields are not sent to the tool API request.
- Outputs: Selected fields
are not returned to the agent.
Note: The quality of the agent may be impacted if critical information is filtered out.
- Block – Prevents tool execution when conditions are met. When using this option, you must provide a reason for blocking. Blocking a tool can result in agent failure.
- Escalate – Assigns an escalation to a user through a compatible escalation app.
After running the agent, the output panel displays:
- The successful execution status.
- The presence of guardrails in the trace.
- The details of applied or skipped guardrails.
The Escalate action requires a compatible action app. A template is available in the UiPath Marketplace under the following name: Agent Tool Guardrail Escalation App.
To make the app available for guardrails, take the following steps:
- Download the app template
from Marketplace and import the
.uipfile into a new Studio project. For details, see Importing app projects in the Studio Web user guide. - You can add new controls and events to the app, but do not change the underlying action schema. Altering the schema renders the app incompatible with guardrail escalations.
- Publish the app, as described in Publishing app projects.
- Go to Orchestrator and deploy
the app in a shared folder, as described in Deploying app projects.
Make sure the folder has serverless robots available, to prevent errors when running the escalation.
- Return to your agent and select the recently deployed app for your guardrails.
