Limitations

Here are the current limitations of Agents in Automation Cloud^TM. These may evolve as the product matures.

Language availability

• Non-English inputs may reduce the quality and reliability of agent runtime. Agents rely on LLMs to interpret inputs and take action through function calling. During runtime, LLMs may handle non-English characters inconsistently, which can sometimes lead to incomplete or inaccurate outputs in the agent loop. If you encounter these issues, we recommend building and training agents in English. Broader language support will depend on future improvements in underlying LLMs.
• Evaluation sets and agent reasoning are displayed in English, regardless of your chosen language.

Tools and activity support

• Process compatibility: Not all process types can be used as tools within Agents. For specific details, refer to the Automations section.
• Not all Integration Service activities are available in Agents. Triggers are not supported, nor activities with mandatory file inputs and outputs. Some generic activities (for example, Get Record) may include attachments or files as available inputs and/or outputs. These fields are not supported in agents. If you need to work with files in your agentic automation, you have the following options:
  • Start with an RPA workflow to fetch a file and pass it as an input to the Run Agent activity.
  • Use an RPA workflow as a tool in your agent to pass a file as an input or output a file to the agent.

Escalations

• Escalations are currently limited to action apps, which generate Action Center tasks.
• Escalations can only be directed to individual users, not user groups.

Other limitations

The following features are not currently supported, but they are planned for future iterations of Agents:

• Versioning: Improved Agent component versioning.
• Permissions: Fine-grained, role-based access control at the agent level beyond what is already supported through Studio Web Solutions RBAC.
• AI Trust Layer: Extensive configuration options for the AI Trust Layer, such as setting restrictions by tool type or action context.