When you create or edit an Agent, three tabs are available. The second tab, AI & Tools, lets you configure the Agent's reasoning behavior and define which additional Agents and Tools it can interact with.
This tab is divided into two main sections:
- AI Configuration
- Tools and Agents
This section defines how the Agent processes and generates responses.
Select the LLM that the Agent will use. Models are organized by provider (e.g., OpenAI, Gemini, Anthropic) and are available in different versions, each with distinct capabilities. Some models prioritize speed, while others are better suited for handling complex requests.
Determines whether consumers can attach files when interacting with the Agent.
When you enable this option, consumers can upload one or more files as long as the total size of all attached files does not exceed 40 Mb.
The supported file types depend on the selected AI Model. After selecting a model, you can view the list of allowed extensions next to this option. For example, this list may include document and image formats.
After you select a model, this setting is populated automatically. It determines how predictable or creative the Agent’s responses will be. Lower values produce more consistent outputs, while higher values allow greater variation.
This value is also automatically filled in after selecting a model. It sets the maximum length of the Agent's responses. Higher values allow more detailed answers.
This setting defines the maximum number of messages that the Agent can send to the LLM during a single execution. Each message counts as a run, including the initial message describing the objective and available resources, as well as any subsequent messages where the Agent updates the context with results from using Tools or Agents. The more resources the Agent uses, the more runs it may require to complete a task. This limit helps control the duration and complexity of each execution.
This setting determines how the AI processes information to generate responses. Depending on the selected strategy, the Agent can produce more logical, creative, or robust answers. The available options include:
-
Enables complex reasoning by guiding the LLM through intermediate steps. This option is ideal for problems that require sequential logic or explanation-based answers. Learn more: Chain of Thought details.
-
Adjusts the prompt in real time based on consumer input or intermediate results, allowing the Agent to adapt its reasoning dynamically as the conversation progresses. Learn more: Dynamic Prompting details.
-
Focuses on improving the consumer’s initial query. The LLM reformulates unclear or ambiguous questions before responding, which leads to more accurate and relevant responses. Learn more: Question Refinement details.
-
Allows the LLM to review its intermediate reasoning, identify potential errors, and revise its output before finalizing the response. This strategy is useful for improving accuracy, but it should be avoided when the expected output must strictly follow a structured format (e.g., a JSON or a list). Learn more: Reflection details.
-
Executes multiple reasoning paths for the same query and selects the most consistent result. This approach increases the reliability of answers in complex or open-ended tasks. Learn more: Self Consistency details.
-
Explores multiple reasoning paths in parallel, maintaining a tree structure of possible thoughts (language sequences). It identifies the best path by iteratively evaluating intermediate ideas. Learn more: Tree of Thought details.
-
Uses the model without prior examples or guidance. The Agent relies entirely on its pre-trained knowledge to respond directly to the input. Learn more: Zero-shot details.
-
Combines zero-shot prompting with explicit instructions for step-by-step reasoning. This approach encourages more deliberate and structured thinking, even for tasks where no examples are provided. Learn more: Zero-shot CoT details.
Guardrails are built-in safety controls you can configure to keep the Agent secure, ethical, and reliable. They help reduce the risk of inappropriate outputs, unsafe consumer inputs, and prompt manipulation.
To configure them, click on Configure Guardrails to open the configuration window. Select the guardrails you want to enable, and then click on Accept to apply the changes, or Cancel to discard them.
-
Helps ensure that the Agent’s responses are appropriate and reliable, reducing biased or unsafe outputs.
-
Protects the Agent from malicious instructions hidden in consumer inputs or system commands that attempt to change its behavior.
-
Detects and manages harmful or inappropriate consumer inputs before the Agent processes them.
This section lets you define which Agents (previously defined and used as Tools within the Agent you are creating) and Tools will be accessible to the Agent you are configuring. In this way, the Agent can execute specific actions, extend its capabilities through external functions, and collaborate with other existing Agents.
To add a new Tool or Agent, click on + Add. A window will open, allowing you to explore all available Tools and Agents.
In the left pane, you'll find a list containing both Tools and Agents.
You can choose to display:
If you're unsure about the purpose or functionality of any Tool or Agent, click on its name to view a detailed description in the right pane.
For example, in the image above, the com.globant.geai.web_scraper_httpx Tool is selected. In the right pane, you can see its name, visibility status (Public), and a brief description of its purpose.
To add a Tool or Agent, select the corresponding checkbox and click on the Save button.