智能体监管

Automation Ops policy settings that govern agents, covering tenant access, guardrails, model availability, and the design limits enforced at publish time.

本节介绍如何使用 Automation Ops™ 策略来管理您组织中的 UiPath 智能体。

Two Automation Ops™ policies govern agents:

• AI Trust Layer policy — Controls tenant-level access to Agents, the guardrails applied to agent LLM calls, and the models agents can use.
• Studio Web policy — Controls the limits an agent must meet before it can be published.

A policy takes effect in the tenants where it is deployed. For the full field-level reference, see Settings for AI Trust Layer Policies and Settings for Studio Web Policies in the Automation Ops user guide.

启用或禁用智能体访问权限

Access to Agents is controlled by the Enable Agents product toggle in the AI Trust Layer policy. The toggle is set to Yes by default.

1. 导航到 Automation Cloud™ 组织中的 Automation Ops™。
2. 为 AI Trust Layer 创建或编辑产品策略。 有关详细信息，请参阅创建监管策略。
3. 在产品开关选项卡上，激活启用智能体选项。
4. 保存策略。
5. 将策略部署到您希望构建或运行智能体的租户。

重要提示：

如果您利用 AI Trust Layer 策略禁用了智能体，已部署该策略的租户中的用户将无法访问 Automation Cloud 中的“智能体”页面。

Guardrails applied to agent LLM calls

Guardrails are configured at two levels: centrally for every agent in a tenant, or individually per agent in Studio Web. For the difference between the two levels, see Guardrails.

The Centralized guardrails tab of the AI Trust Layer policy defines the guardrails that apply to every agent of a covered type in the tenant. Detection settings are configured once per guardrail type and applied uniformly to every agent. For the configuration options of each type, see Centralized guardrails.

备注：

The Enable LLM as Judge guardrail product toggle governs the LLM as Judge guardrail separately. Because that guardrail makes its own LLM call, the call is also subject to the centralized guardrails configured in the policy.

Models available to agents

The Models tab of the AI Trust Layer policy controls which UiPath-managed models agents can call. Agents and coded agents each have their own product-level model controls. For example, a policy can restrict agents to Bring Your Own Model configurations, turn off a model provider, or lock agents to the currently selected model versions so that newly released models are not allowed automatically. For the full list of controls, see Models.

These controls apply to UiPath-managed models only. Models configured through a custom LLM configuration remain available regardless of how the Models tab is set.

The Model Regions tab restricts the geographic regions that LLM traffic is routed to, including the calls agents make. Routing follows the default region configuration unless it is set to Custom, in which case only the selected regions are used.

Other AI Trust Layer toggles that affect agents

Beyond Enable Agents, the Product Toggles tab carries other toggles that change what agents can do. For example, disabling calls to third-party AI models blocks those calls for every product that uses them, including agents, and disabling Context Grounding causes agents that depend on it to fail. For the full list of toggles and their defaults, see Product Toggles.

Agent design settings in the Studio Web policy

The Design tab of the Studio Web policy sets limits that an agent must meet before it can be published. The limits are enforced in the agent builder in Studio Web. For example, a policy can require a minimum agent score, cap the temperature an agent can be configured with, or require the agent to define at least one Human In the Loop escalation.

These limits apply to new agents and to new versions of existing agents. An agent that is already published continues to run unchanged, but an updated version of it must meet the limits before it can be published again.

备注：

If the policy sets a maximum number of tokens per agent response, the effective value can vary by model. The value shown in the agent builder follows either this policy or the model's maximum token limit.

For the configuration options of each setting, see Design in the Automation Ops user guide.