Building AI Agents with UiPath: A Step-by-Step Guide

UiPath Studio Web categorizes agents into autonomous, conversational, and coded types, each serving distinct automation needs. The autonomous agents are designed for automated workflows using the autopilot option.

Conversational agents function as chat-based interfaces for interactive dialogue, while coded agents cater to developer-focused automation tasks. This tutorial focuses primarily on the autonomous and conversational agent types.

The prompt restricts the agent's focus to its specific purpose, such as generating email replies.

User prompts, in contrast, serve as the dynamic input that the model processes for each specific task. This distinction ensures the agent adheres to its defined role while handling varied user requests.

Variables are frequently used within prompts to inject external data, managed through the data manager, allowing for flexible and dynamic agent interactions.

By selecting the plus sign within the data manager tab, developers can configure input variables and mark them as required fields to ensure dynamic interaction. Developers can create new variables, such as 'user input,' and provide a clear description of their purpose.

These variables can be marked as required fields, ensuring that agents prompt for necessary information before execution. Variables are inserted into prompts using double braces, for example, `{{user input}}`.

During the debugging phase, the agent triggers an input request for these variables, allowing for real-time testing and validation of the dynamic inputs.

AI agents are conceptualized as the 'brain' of the operation, responsible for thinking, interpreting, and making decisions. This analogy highlights their role in processing information and determining appropriate actions.

Tools are likened to the 'hands' of the agent, executing the actions decided upon by the AI brain. This separation of function ensures a clear division between decision-making and task implementation within the automation process.

This connection enables agents to perform email-related tasks.

The UiPath Orchestrator plays a vital role by managing connection verification and storage folders, ensuring seamless communication between the agent and the email system.

Input arguments can be defined in the Data Manager to pass agent output directly into the workflow body, enabling the agent to dynamically populate email content based on its processing.

An AI agent may generate the correct output but fail to invoke an external tool if specific instructions are missing from its prompt. The agent requires explicit cues to route data to a designated tool.

These cues must be provided either in the system prompt or the user prompt to guide the agent in identifying and utilizing the correct utility. Without clear instructions, the agent may not recognize when to trigger an external action.

Renaming tools with descriptive names and providing precise descriptions are essential for the large language model (LLM) to accurately identify and use the appropriate tool for a given task. This clarity prevents misinterpretations and ensures successful tool execution.

For example, if an agent is meant to send an email, the tool name and description must clearly indicate its email-sending functionality, alongside a prompt instructing the agent to 'send an email to X with the following content'.

To prevent formatting issues and ensure consistent output, agents must be strictly instructed on response formatting, avoiding symbols like hash or star. Instead, the agent is mandated to use specific HTML tags for structuring its responses.

This approach ensures that paragraph collapsing is prevented and the output maintains a predefined structure. An example template provided includes headers and bullet points, all defined using HTML tags.

By enforcing HTML formatting, developers can guarantee that the agent's output is consistently readable and aligns with desired presentation standards across different applications.

Because raw Word or PDF files remain inaccessible during execution, the system relies on a digital index to feed readable data to the agent. Instead, they interact with an index that translates file data into a digital format the agent can process.

This index acts as the primary feed for providing data to the agent, enabling it to read and understand the content effectively. The system relies on this indexed data to perform its operations.

Therefore, creating and maintaining accurate indexes is crucial for ensuring that agents have access to the relevant information required for their tasks, optimizing their performance and accuracy.

Initial email outputs contained repeated greetings, such as 'hi team' and 'thanks AI agent,' due to duplication within the workflow design. This redundancy led to unprofessional and cluttered emails.

To resolve this, the redundant lines were directly removed within the workflow editor, streamlining the email generation process. This ensures that the agent's output is concise and professional.

The fix was confirmed by rerunning the agent, which then produced a clean email output without any repetitive text, demonstrating the importance of careful workflow design.

Removing indices and storage buckets allows an agent to execute tasks like email generation by relying exclusively on updated system instructions. The developer demonstrated this by removing indices and storage buckets from the agent's configuration.

The agent then relied solely on the updated system prompt to perform its task, proving that direct instructions can be sufficient. This simplifies the agent's setup for certain applications.

The agent successfully executed the email generation task using only the revised system instructions, confirming the viability of this approach. This flexibility allows for optimizing agent configuration based on task complexity.

This method involves carefully crafting the system prompt to include all necessary information that would otherwise be provided through context grounding, ensuring the agent remains focused and accurate.

Autonomous agents are primarily designed for specific task completion, operating on a simple input-processing-output model without extensive interaction. Their goal is to efficiently execute predefined tasks.

Conversational agents, however, function as digital workers specialized in multi-turn dialogues, engaging in multiple rounds of discussion with users. They are built for interactive and dynamic communication.

Creating conversational agents in UiPath Studio requires selecting a specific agent type during setup, tailored to facilitate extended user interactions. This initial configuration ensures the agent's capabilities align with its communicative role.

For a new conversational agent, an initial HR policy assistant prompt is set up to guide its responses, establishing its persona and knowledge domain for engaging with users on HR-related queries.

All automations should be created in a shared folder to ensure accessibility across different parts of the system.

PDF documents containing essential information, such as holiday policies, PTO guidelines, and travel expenses, are uploaded to these storage buckets. This centralizes the data for agent access.

A 'Document Index' is then created within the storage bucket, allowing the agent to efficiently query and retrieve relevant data from the uploaded documents. The 'basic' index setting is recommended for text-based documents to optimize costs and performance.

This setup ensures that the conversational agent has a robust and organized knowledge base to draw upon when responding to user inquiries regarding HR policies.

The document index created earlier is linked to the conversational agent via the 'context grounding' feature, providing the agent with a knowledge base for its responses. This connection is crucial for ensuring accurate and relevant information retrieval.

The 'Retrieve' result count for the index is adjusted to five or ten to optimize data retrieval accuracy, ensuring the agent fetches the most pertinent information. This fine-tuning enhances the agent's ability to respond effectively.

An 'API Workflow' tool is added to the agent to enable email sending capabilities, allowing the agent to perform external actions such as sending notifications. This expands the agent's functionality beyond conversational responses.

A new connection to Microsoft Outlook is established within the shared folder to facilitate automated email interactions. This integration ensures seamless communication between the agent and external email systems.

This mechanism ensures transparency in agent operations.

The output configuration is specifically set to report the status of the email sending task, providing immediate feedback on whether the action was completed successfully. This is crucial for debugging and operational monitoring.

Testing confirmed that the workflow successfully completes after integrating this response logic, validating the effectiveness of the automated feedback system. This enhancement improves the overall reliability and user experience of the agent.

Initial tests revealed that generated emails were in plain text, deviating from the desired HTML format. This issue necessitated a refinement in the agent's instructions to ensure consistent formatting.

To address this, a specific instruction was added to the system prompt: 'always send the email in HTML format'. This explicit directive guides the agent's output behavior.

Subsequent testing confirmed that the agent correctly followed this new directive, generating emails in the required HTML format. This demonstrates the power of prompt engineering in controlling agent output.

This refinement ensures that all email communications generated by the agent adhere to professional formatting standards, improving readability and presentation.

The UiPath Orchestrator's agents tab functions as the central hub for monitoring, where users verify that the cloud-based serverless instance successfully executes tasks. It successfully operates on a cloud serverless instance, demonstrating its scalability and deployment flexibility.

User testing confirms that the agent retains its ability to generate content and execute email tasks accurately after deployment. This verification step is crucial for ensuring the agent functions as intended in a live environment.