Built-in AI capabilities

Every AI Agent gambit can use a set of native tools. You do not connect a toolkit or write a custom action to use them. Each tool starts off and is enabled per gambit, so one agent can search the web while another cannot.

​

The Built-in Tools & Capabilities dialog

All native tools are enabled in one place. Select an AI Agent gambit in the builder, then select Configure in the Tools & Capabilities region of its editor. The Built-in Tools & Capabilities dialog opens. The dialog groups everything into three sections. The Built-in Tools section holds the five tool cards described below. Two more sections hold the interface components the agent may render. Each section header names its underlying tool: Output components (render_ui) and Input components (request_user_input). A search field at the top filters across all sections, and Save applies your selection. Enabled tools then appear as badges in the gambit editor. This page covers only this dialog and its follow-up settings. For the rest of the gambit editor, see AI Agent gambit.

​

Web Search

The Web Search card lets the agent search the internet for real-time information during a turn. Enable it when the agent needs answers that are not in your knowledge bases.

​

Code Interpreter

The Code Interpreter card lets the agent execute code to perform calculations and data analysis. The agent runs code when a question needs real computation, which keeps numeric answers exact.

​

File Generator

The File Generator card lets the agent generate files such as PDFs, CSVs, and images. The agent returns each generated file inside the conversation, where the end user can open it.

​

Data Extraction

The Data Extraction card lets the agent pull typed fields out of conversation text. It is the only card with follow-up configuration. When it is on, a Data Extraction section appears in the gambit editor, where you define the extraction schema one field per row. Each field has a name, a type, and a description that tells the agent what to capture. A Required switch marks the field as required. The Save to flow variable input stores the extracted value in a flow variable, so gambits later in the flow receive a clean value instead of prose.

​

File Reading

The File Reading card lets the agent read files the end user sends during the conversation. When it is on and the conversation holds at least one upload, the agent gains two tools, read_document and get_document_index.

​

What the agent can read

Documents are converted to text before the agent sees them. Supported formats are PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and plain text such as TXT, Markdown, JSON, XML, and YAML. An image is indexed as a description of what it shows plus any legible text in it. A scanned PDF with no text layer is transcribed the same way, and video and audio files are not read. Uploads reach the agent the same way from every entry point. The entry points are the Attach files button in the composer, a file upload input component, and a Media Upload gambit earlier in the flow. A document or image sent on WhatsApp is read the same way.

​

How a read works

The agent never receives a whole file. Each turn, its prompt lists every uploaded file with a status, a line count, and up to eight labelled sections. The agent reads only the sections it needs, up to 8,000 characters per call, and continues from a cursor when a read is cut short. For a file with more sections, get_document_index returns the full list. Two uploads with the same name are told apart by a number. A second photo.jpg is listed as photo (2).jpg, so a read always lands on the intended file. When Show tool calls to visitors is on, each read appears in the chat like any other tool call.

​

Status and limits

A file is listed as still being read, ready, could not be read, or unsupported. A read of a file that is still being read waits up to 20 seconds. After that, the agent is told the file is still being read and to ask the end user for a moment.

​

Interface components

Each of the two sections has its own switch, and you select individual component cards to control exactly what the agent may render. The book icon on a section or a card opens its documentation inside the dialog. Output components render alongside the agent’s reply. They include cards, data tables, charts, timelines, image galleries, and metric displays. Input components render an interactive element and wait for the end user’s response. They include forms, button groups, calendars, time slots, file upload, and one-time-code verification.

​

Knowledge citations

Citations are not a card in this dialog. When an agent answers from an attached knowledge base, the sources behind the answer can be shown to end users as citations. See Attach knowledge to an agent for how knowledge reaches the agent.

​

What end users see

Visibility of all this activity is controlled per agent, under Debug in the Settings list on the Configure page, not in the builder. Hiding activity changes only the presentation, not the agent’s behavior. For the three switches and their defaults, see Test your agent. For what reasoning, tool calls, and citations look like in the chat, see End-user experience.

​

Related pages

• AI Agent gambit
• Guardrails and safety
• Attach knowledge to an agent
• Media Upload gambit
• End-user experience