- Démarrage
- Agents conversationnels UiPath
- Agents UiPath dans Studio Web
- À propos des agents UiPath
- Licences
- Agents avancés (aperçu)
- Exécuter des agents
- Agents et workflows
- Meilleures pratiques pour créer des agents
- Choix du meilleur modèle pour votre agent
- Meilleures pratiques pour la publication et le déploiement d’agents
- Meilleures pratiques en matière d’ingénierie contextuelle
- Meilleures pratiques pour DeepRAG et la transformation par lots : JIT vs. stratégies basées sur l’index
- Utiliser des fichiers
- Contextes
- Escalades
- Évaluations
- Traçages d’agent
- Examen de l’agent (aperçu)
- Agents codés UiPath
- Créer avec Coding Agents
Gestion des garde-fous
Runtime guardrail coverage, checks, detections, and applied actions across your agents on the Agents page, with drill-down into individual agents and traces.
The Guardrails tab shows how the guardrails configured for your agents behave at runtime. It brings together what guardrails checked and detected in the selected period, and which agents and execution stages those checks covered.
For details on adding guardrails to your agents, refer to Guardrails.
The tab is split into an Overview and a Coverage view.
From the tab, you can:
- See how many active agents had at least one guardrail check, and how many agent executions guardrails blocked.
- Compare checks and detections by guardrail type and by policy source.
- Find agents that ran without guardrail checks, or LLM calls that both policy sources checked for the same guardrail type.
- Open an individual agent and its traces to investigate a detection.
The data covers the selected time range, which defaults to the last 30 days.
Checks, detections, and policy sources
The following labels appear throughout the Guardrails tab:
| Label | Signification |
|---|---|
| Vérifier (Check) | One evaluation of a guardrail at one execution stage. |
| Detection | A check in which the guardrail found what it is configured to detect, such as personally identifiable information (PII). |
| Detection rate | The share of checks that resulted in a detection. |
| Étape de l’exécution | When a check ran: before (Pre) or after (Post) an Agent turn, an LLM call, or a Tool call. |
| Agent-defined | The check came from a guardrail configured on the agent itself. |
| Centralized | The check came from a centralized guardrail enforced through an AI Trust Layer policy. |
| Both layers | The same stage was checked by both policy sources, an agent-defined and a centralized guardrail. |
| BYO | Bring your own. The guardrail runs on your organization's own subscription to a supported safety vendor. |
The execution stages correspond to the agent, LLM, and tool scopes you select when configuring a guardrail. The actions a guardrail can apply after a detection, such as Block or Log, are described in Out-of-the-box guardrails.
Vue d'ensemble (Overview)
The Overview view summarizes guardrail activity across your agents for the selected time range.
The view contains the following elements:
| Élément | Ce qu'il affiche |
|---|---|
| Agents checked by guardrails | How many of the active agents in the period had at least one guardrail check. |
| Agent executions blocked by guardrails | How many agent executions contained a Block action, and their share of the agent executions in the period. |
| Guardrail types over time | Checks or detections for each guardrail type, grouped into time buckets. Checks and Detections switch between the two counts. |
| Guardrail check composition | The check volume for each guardrail type, split by policy source, with the total checks and detection rate for that type. |
| Actions applied over time | The Block, Escalate, Filter, and Log actions that guardrails applied, grouped into the same time buckets. |
Couverture
The Coverage view shows which execution stages guardrails checked, for each agent and for each guardrail type. It also lists patterns in the runtime data that may need attention.
Agent coverage
The Agent coverage table lists every agent with runtime activity in the selected period. For each agent, the table shows:
- Agent executions, Checks, and Detections – The totals for the period.
- Agent turn, LLM call, and Tool call – One cell for each stage, Pre and Post, with the number of checks that ran there.
Checks and Detections switch the stage cells between the two counts. You can search the table by agent name and sort it, for example by agent executions or by checks. Selecting an agent opens its agent guardrail details.
Guardrails coverage
The Guardrails coverage table shows the same stage breakdown for each guardrail type, across all agents. Under each guardrail type, the table names the layers that checked it, and a BYO label marks types that include bring-your-own guardrails. Checks and Detections switch the table between the two counts.
Problèmes potentiels
The Potential issues list points out patterns noticed in the runtime data for the selected period. Whether a pattern needs action depends on what each agent is meant to do. Examples include:
- Agents that ran without a single guardrail check – The issue links to Agent coverage, sorted by checks from lowest to highest, so the agents without checks appear first.
- LLM calls checked by both layers for the same guardrail type – A centralized policy and an agent-defined guardrail evaluated the same call. Expanding the issue lists the affected agents.
Agent guardrail details
Selecting an agent in Agent coverage opens its guardrail details. The following sections show guardrail activity for the version selected in Agent version. By default, this covers all agent versions observed in the selected period. Selecting a single version narrows every section to that version, for example to investigate an issue in a specific release.
Summary and trends
The view opens with two summary cards:
- Checks – The total number of checks for the agent, with the number of evaluated guardrail types.
- Agent executions blocked by guardrails – The share of the agent executions that guardrails blocked, with the number of blocked executions and a via escalation count.
The Guardrail types over time chart that follows the cards shows the checks or detections for each guardrail type of this agent, in the same time buckets as the Overview view.
Guardrail checks
The Guardrail checks table lists each guardrail evaluated for the agent and the selected version.
| Colonne | Description |
|---|---|
| Garde-fou | The guardrail name and its guardrail type. A BYO label marks guardrails that run on your organization's own safety vendor subscription. |
| Source | Agent-defined or Centralized. |
| Scope / stage | The scopes the guardrail checked, such as Agent or LLM, and its execution stages, such as Pre & post-execution. |
| Vérifications | The number of checks the guardrail ran. |
| Detections | The number of checks that resulted in a detection. |
| Detection rate | The share of checks that resulted in a detection. |
| Latest detection | When the guardrail last detected something, or None in period. |
Traçages d’agent
The Agent traces table lists the complete traces in which the agent had matching guardrail activity. You can narrow the list by:
- Trace ID
- Guardrail types
- Policy source
- Actions applied
Filtering by Trace ID requires the complete trace ID.
You can also limit the list to traces that match one or more of the following options:
- Has detections
- Blocked by guardrails
- Has BYO guardrails
Sort traces orders the results, with the newest traces first by default. Apply applies the selected filters, and Reset clears them.
| Colonne | Description |
|---|---|
| Trace ID | The trace identifier, with an option to copy it. |
| Démarré (Started) | When the trace started. |
| Vérifications | The number of guardrail checks in the trace. |
| Detections | The number of detections in the trace. |
| Actions applied | The actions guardrails applied in the trace and how many times, or None. |
| Garde-fous | The guardrail types and policy sources involved in the trace. |
| Duration | How long the trace ran. |
| État (Status) | Whether guardrails blocked the trace: Blocked or Not blocked. |
| Détails du traçage | View trace opens the full trace. |
Détails du traçage
View trace opens the trace in the Trace details panel. In the Execution Trace tree, guardrail checks appear as steps at the stage where they ran, for example:
- Agent input guardrail check – The agent-defined guardrails that evaluated the agent input.
- LLM input guardrail check – The agent-defined guardrails that evaluated the request to the model.
- LLM Request Centralized Guardrail Check and LLM Response Centralized Guardrail Check – The centralized guardrails that evaluated the model request and response.
- Block action – The block applied after a detection.
Selecting a guardrail step shows its raw JSON, including the validation result and the details of what the guardrail detected. For more information on reading traces, refer to Agent traces.
Investigating a guardrail detection
To find out why a guardrail triggered, follow the detection from the Guardrails tab to the step in the run where it happened.
- In the Guardrails tab, select Coverage.
- In Agent coverage, select the agent to investigate.
- Optional: Choose a version from Agent version to narrow the view to that version.
- In Agent traces, select Has detections.
- Sélectionnez Appliquer.
- For the trace to investigate, select View trace.
- In the Execution Trace tree, select the guardrail step to inspect, for example PII Detection.
The step details show the raw JSON for the check, including the validation result and what the guardrail detected.