This is a step-by-step guide to create, test, and publish an Agent.
First, log in to the Console. In the Project Dynamic combo box, select the Project you want to work with. In this case, Documentation (Demo) is used.
Next, on the left side of the screen, locate the Backoffice menu. In this menu, click on The Lab.
A new browser window will open with The Lab.
Once inside The Lab, the Agents Dashboard opens by default. From there, you can create an Agent using one of the following three methods:
- New Agent with Iris: For a more intuitive and streamlined experience, The Lab provides an assisted creation method guided by Iris, the integrated Assistant. Describe what you want the Agent to do, and Iris will help you configure key fields such as AI Model, Reasoning Strategy, and appropriate Tools.
- Import Agent: Import a LabPackage file, which can include an Agent and its resources (Tools or other Agents).
- New Agent Manually: Full control over your Agent’s configuration. Define specific tasks, select appropriate Tools, and fine-tune the Agent's behavior to precisely match your requirements.
To begin, select one of the options and click on the corresponding button or link:
- In the New Agent with Iris card, click on the Start with Iris button.
- In the Import Agent card, click on the Import Agent button.
- Under New Agent Manually, click on the Start from scratch link.
If you selected Import Agent in Step 2, the LabPackage already includes the Agent’s purpose. Otherwise, define the purpose using one of the following options:
- 3.1. New Agent with Iris
- 3.2. New Agent Manually
Selecting New Agent with Iris automatically opens a chat window with Iris, the AI Assistant, as shown below:
At this stage, you can provide Iris with details about your Agent via chat, including:
- Name: Give your Agent a descriptive name (e.g., "Customer Support Agent" or "Data Analysis Assistant").
- Role: Describe the primary function of your Agent (e.g., "Assist customers with product inquiries" or "Analyze sales data to identify trends").
- Purpose: Explain the overall goal you want your Agent to achieve (e.g., "Improve customer satisfaction" or "Provide data-driven insights to the sales team").
Suppose you want to create a marketing research Agent. In the chat window, you could type:
"I want to create an agent named 'Marketing Research Assistant.' Its role is to analyze market trends and competitor activities in the technology industry. The purpose is to provide the marketing team with data-driven insights to inform our marketing strategies and campaigns."
Once you've provided Iris with this information, press Enter or click on the arrow button to send your message.
Selecting New Agent Manually opens the Configuration tab, where you can define essential details about your Agent.
- Name: For example, "Marketing Research Assistant".
- Role: Specify it clearly, such as "Analyzes market trends and competitor activities in the technology industry".
- Purpose: Clearly outline your Agent's objective, such as "Provide the marketing team with data-driven insights to inform marketing strategies and campaigns".
In this example, the same Agent described in step 3.1 is defined, where Iris was instructed to "create a Marketing Research Assistant".
Note that even when choosing the manual creation method, a chat panel remains available on the left side of the screen. You can still interact with Iris in this panel to receive suggestions or assistance while configuring your Agent. You can describe the Agent you want to build, ask questions, or request help completing certain fields, making the manual process more flexible and guided if needed.
Take a moment to review, complete, and finalize your configuration depending on your creation method:
- 4.1. New Agent with Iris: Review Agent Details
- 4.2. Import Agent: Review and Customize
- 4.3. New Agent Manually: Configure and Customize
If you created your Agent using Iris, the system will automatically fill in all the necessary details for your Agent based on your conversation with Iris.
Review the following:
- Configuration tab: This tab focuses on defining your Agent's identity and behavior. Here you should check:
- Name, Role, and Purpose: Verify that these core elements accurately reflect your instructions to Iris.
- Avatar Image: Optionally, you can generate an avatar image based on the Agent’s name and description. To do this, click on Generate image.
- Background Knowledge and Guidelines: Iris can automatically fill in these fields with suggested data and rules. Review and adjust them as needed. Keep in mind that at least one of these two fields must be completed before publishing the Agent.
- Examples: Iris may provide examples of how to interact with your Agent. Review these to ensure they align with your expectations.
- AI & Tools tab: While Iris handles the initial configuration, it's good practice to review the AI & Tools tab. Here, you can check the selected AI Model, Reasoning Strategy, and enabled Tools.
- JSON tab: Automatically reflects all configuration changes you make, providing a structured view of your Agent's setup.
Although Iris takes care of the initial setup, you retain full control over your Agent's configuration. You can make adjustments in two ways:
- Direct editing: Modify settings directly within each tab. You can experiment with different AI Models and Reasoning Strategies or adjust the enabled Tools to optimize your Agent's performance.
- Iris Guidance: If you are unsure about a specific setting or need assistance, Iris is ready to help. In the chat panel on the left, describe the changes you want to make or ask specific questions. Iris can guide you through the options, explain the implications of different settings, and even make the adjustments for you within the configuration.
When importing a LabPackage, The Lab automatically completes all the Agent's data. However, it is advisable to review the three tabs and adjust any details as necessary:
- Configuration: Validates the Name, Role, Purpose, and other details.
- AI & Tools: Checks the AI Model, Reasoning Strategy, and associated Tools.
- JSON: Reviews the complete Agent definition with the settings you have configured.
If you have created your Agent Manually, you have full control over its configuration. Make sure to configure the following:
- Generate Avatar Image (Optional): You may generate an avatar image for your Agent.
- Background Knowledge and Guidelines: Provide any relevant context or specific instructions to guide your Agent's behavior. You must complete at least one of these two fields, as they inform the Agent what to do and how to respond. If both fields are empty, the Agent cannot be published.
For this case, it is useful to combine context and response rules, so you can fill in the following information:
- Background Knowledge: You are a Marketing Research Assistant, specializing in analyzing market trends and competitor activities in the technology industry. Your goal is to provide the marketing team with data-driven insights to guide their strategies and campaigns.
- Guidelines:
- Analyze the information in two sections: Market Trends and Competitive Analysis.
- Include key figures (market size, CAGR, share) whenever available.
- Mention the source or year of the data in parentheses.
- Summarize the findings in no more than 200 words.
- If data is missing, indicate "Data not available" instead of speculating.
- Maintain a professional and concise tone.
- Provide Examples (Optional): Add a few examples of how you intend to interact with the Agent. This helps refine its understanding of your requests.
For example, you can add the following information:
- Input Data: What are the current trends in AI technology?
- Output Data: Market trends: Adoption of generative models in healthcare and finance is growing; global investment in AI is projected to reach $200 billion by 2025 (CAGR 18%). Competitive analysis: OpenAI leads with GPT-4o; Google promotes Gemini; Anthropic focuses on model alignment.
-
Configure AI & Tools tab: Navigate to the AI & Tools tab. Select a suitable AI Model (for example, gpt-4o) and Reasoning Strategy (for example, Chain of Thought) that align with your Agent's purpose. Consider the complexity of the tasks you want your Agent to perform and choose a model that can handle them effectively.
If you want consumers to upload files when interacting with this Agent, make sure you enable Allow users to upload files (jpeg, jpg, gif, png, pdf, txt, doc, docx, odt, rtf, ppt, pptx up to 40 MB).
In the Tools and Agents section, enable only the Tools that your Agent requires. For the Marketing Research Assistant, make sure that the following Public Tools are enabled:
- com.globant.geai.web_scraper_httpx
- com.globant.geai.web_search
These tools are enabled by default in The Lab and allow the Agent to efficiently extract and analyze online data.
If you created your Agent using Iris or imported it, The Lab will automatically fill in all the data. After confirming the settings, click on the Run Test button in the lower-right corner of the screen to test your Agent.
If you created your Agent manually, after clicking on the Create Agent button, an on-screen confirmation message will appear indicating that your Agent has been successfully created. At the bottom right, the Create Agent button changes to Save, and the Publish to The Station and Run Test buttons become available. To start testing your Agent, click on Run Test.
In any case, when you click on Run Test, a chat window opens on the right side of your screen. Here, you can interact directly with your Agent, test its capabilities, evaluate its responses, and ensure it operates as intended. If you click on the attachment icon below the message box, you can also upload content as a File, Image, Video, or Audio for the Agent to use during the test.
If you're ever unsure about what your Agent is doing or encounter issues, you can click on the Debug tab.
For each message you send, a sequence of steps is recorded:
- PopulateLLMRequest
- FunctionTable
- initial-message
These allow you to review the data that the Agent sends to the model, view the functions available to the Agent, and check the initial messages sent to the model.
When the Agent invokes a Tool, Debug adds a block per call:
- function-call:<Tool_name> – The request that the Agent sends to the Tool.
- message – The response that the Tool returns.
In this case, a single message was sent, which generated the base steps (PopulateLLMRequest, FunctionTable, initial-message) and three Tool invocations: one to com.globant.geai.web_search and two to com.globant.geai.web_scraper_httpx. Each function-call is followed by its message block with the Tool's response.
In the top bar, you will find the following:
-
Enables or disables the preview of each step the Agent executes while responding. This option is enabled by default.
-
Allows you to download a JSON file containing the complete debug log (messages, model calls, and metadata) for the current conversation.
-
Opens the options menu, where you can select Clear conversation to delete the current chat and start a new one.
-
Closes the chat window.
When you have adjusted your Agent and are ready to test it, follow these steps:
- At the top of the chat, you will see "Reset conversation to test latest changes". Click there to start a fresh conversation and validate the new settings (you can also choose ... > Clear conversation).
- If something goes wrong, make the necessary changes and restart the conversation before testing again.
When a consumer interacts with an Agent from the Workspace, they can share the conversation they are having with the Agent via a link. Anyone who accesses that link can view the conversation history up to the shared point and continue it from there.
To enable this feature, click on the three-dots menu in the upper-right corner of the Agent and select Chat sharing settings. A pop-up window will open with the following options:
- None: The conversation cannot be shared via a link.
- Anonymous: Allows the conversation to be shared via a public link.
You must select Anonymous to allow a consumer to share their conversation. In this case, anyone with the link will be able to access the conversation under the following conditions:
- Access is anonymous and does not require login.
- The shared conversation expires based on time and/or number of interactions.
- Interactions made from the link are not saved in the history. If the page is reloaded or closed, the conversation progress will be lost. The original shared history will remain available as long as the link is valid.
More information: Sharing a conversation with an Agent with other consumers via a link.
You can configure how the Agent will be introduced to consumers once it is published in the Workspace. The presentation defines what consumers see before starting the interaction and helps set the right context.
In Introduction, you can add a short welcome message, such as "Your marketing insights assistant is ready to help".
In Description, include a brief explanation of the Agent’s purpose; for example, "Provides data-driven analysis of market trends and competitor activities to support business decisions".
In Conversation Starters, you can suggest initial questions to guide the consumer. For example:
- Starter Title: Market trends
- Starter Prompt: What are the latest market trends in the technology industry?
In Features, highlight the key strengths of the Agent. For example:
- Feature Title: Market trend analysis
- Feature Description: Generates clear and concise insights on current trends supported by data.
For more details on each field and configuration option, see Configuration Tab of an Agent.
Once you have tested your Agent and verified that it works correctly, you can proceed with saving it. To do so, go to the bottom-right corner of the screen and click on the Save button.
Saving the Agent makes it available in the Workspace as an Assistant.
If you selected Anonymous in Step 6: Set up the option to share conversations, the Share option becomes available after the Agent replies at least once.
The Share icon appears in the chat header.
Click on Share to open the Share Conversation pop-up window with a public URL and a Copy link button. The link does not expose your identity or custom instructions. The recipient can view the conversation up to the shared point and continue from there. Messages you send after sharing in your original chat are not included in the link.
Anyone who opens the link will access the chat as Guest. The header displays Guest and the label Continuation of a shared conversation.
The guest can view the history up to the shared point and continue the conversation from there.
The guest cannot start a new chat with the Agent from that link. They can only continue the shared conversation.
To configure who can execute your Agent after it is published, click the three-dots menu in the upper-right corner of the Agent editor and select Execution permissions. In the window that opens, choose the option that best fits your case (for example, None or Project). The options available depend on how permissions are configured, as described in Publishing an Agent to The Station. Click Save to confirm.
Once execution permissions are set, go to the bottom-right corner of the screen and click on Publish to The Station to make the Agent available from The Station.