Building an agent

This section describes how to create an agent. Once you understand the basics, check out Best practices.

1. In UiPath® Automation Cloud^TM, select Agents from the left-hand navigation bar.
   The Agents workspace is displayed.
2. You can create your own agent, import one from a JSON definition, or start from one of the available templates. For details, refer to Using prebuilt agents.
3. To create an agent from scratch, select New agent.
4. The Agents workspace includes three tabs: Design, Evaluations, and Traces.

   Use the Design tab to design and define the core elements of your agent. Once you define the input/output arguments, you can use this space to preview your agent runs.

   Use the Evaluations tab to create evaluations and store results. Evaluations objectively measure your agent's output and assess whether or not it is consistent with your objectives.

   Use the Traces tab to log and represent all design and evaluation time traces for this agent. This is where you can see all the agent runs and the details of the run.

5. First, select the large language model you want to use with your agent. Models are provisioned through UiPath AI Trust Layer.
   1. To configure model settings, select the settings icon next to the Model field. Here you can configure the Temperature and Max. tokens per response fields.
      • The Temperature determines the creativity factor of the LLM generation. You can set a value between 0 (Precise) and 2 (Creative).
      • Max. tokens per response refers to the maximum number of tokens to generate with the agent's response.
6. In the General tab, give your agent a unique name and description.

   The name and description of the agent must offer details regarding the purpose of the agent in simple language. The description is used to help identify the agent in Studio.

7. In the Prompt tab, add a well-structured system prompt and a user prompt to give your agent instructions for what it should do, and constraints it should follow.
   For details, refer to System prompts and User prompts.
8. Select Arguments to access the Arguments window, where you can configure the input and output arguments used in your prompt. For details, refer to Prompt arguments.
9. In the Tools tab, add any tools your agent can use during execution. These can be a set of Integration Service activities and published processes to which you have access.

   Each tool must have a detailed description that helps inform the agent how and when to use this tool. For details, refer to Tools.

10. Use the Contexts tab to add any necessary context.

    Context Grounding indexes give an agent access to permissioned knowledge bases. This helps ground agent actions or answers in your business data. For details, refer to Contexts.

11. Use the Escalations tab to add a human in the loop and enable Agent Memory.

    Adding an escalation allows your agent to involve a human to review, approve, or update output as it is running. For details, refer to Escalations and Agent Memory.

12. In the Playground window, provide input for a test run and select Run.

    You must test your agent to ensure it's operating as expected.

13. Next, go to the Evaluations tab to review and measure your agent.

    An evaluation is a pair between an input and an assertion made on the output, and can be used to grade your agent during an end-to-end run. For details, refer to Evaluations.

14. After you run the evaluations, use the Traces tab to review all the trace logs for your agent. For details, refer to Traces.
15. Once you have tested, evaluated, and refined your agent, select Publish to publish it to Orchestrator as a process.
    • Select where to publish:
      • Orchestrator Personal Workspace Feed - The process is published to your personal workspace and will only be available to you. This is the default selection.
      • Orchestrator Tenant Processes Feed- The process is published to the global tenant feed and will be available to all users assigned to folders that use the tenant feed.
    • Select a version. A version is generated automatically starting with 1.0.0 when you first publish, and then incremented automatically every time you republish.

    Important: This step is necessary for the agent to be available in the Run Agent activity.

    Select Export to view and package the agent definition as a JSON. You can then use this JSON extract to import the agent in a different organization.

16. Finally, insert your agent into an automation using the Run Agent activity in Studio Web or Studio Desktop. For details, refer to Running agents.