Sigma agents (Beta)

You can add Sigma agents to workbooks to provide AI interfaces to users in Sigma. You can add agents specific to business tasks, like forecasting different sales promotions, or summarizing the status of an incident in a debugging app. Users can interact with agents that you build using natural language in a conversational chat interface.

Sigma agents are specific to a workbook, based on the predefined context of the data elements in the workbook and the access granted on the workbook. Instructions and data sources provide the context, while actions and agents configured in your data platform provide the tools for an agent to take action or enrich insights.

An AI agent is not a large language model (LLM), but rather uses LLMs in combination with context and tools like other warehouse agents, MCP servers, and actions in Sigma. AI models generate probabilistic outputs, but you can ensure consistent execution of critical tasks by providing deterministic action tools to your agent. Action tools run pre-defined actions, ensuring that the actions run consistently and predictably every time. The agent uses probability-based reasoning to decide when to use a tool.

Example use cases

For example, you might build an agent to address one of the following use cases:

• Automated forecasting and review workflow: An agent can run forecasting calculations based on the contents of an input table and informed by policy documents in an external system, then present results in a structured format for review and approval.
• Personalized sales copilot: An agent can review pipeline opportunities that a specific user is working on, flagging customers with upcoming renewal dates and drafting Slack message follow-ups that the user can send to specific channels.
• Analyze key drivers of customer satisfaction: Based on customer satisfaction survey data, an agent can identify, explain, and rank the drivers that contribute to low or high customer satisfaction scores. For specific customers represented in the survey data, the agent can send Slack or Microsoft Teams messages, such as to notify customer success managers about accounts that might be a churn risk.
• Detect anomalies in purchasing behavior: An agent can review sales data across customer segments and identify unusual transaction patterns, based on both deterministic guidelines that you provide and probabalistic reasoning for newly identified patterns. After detecting an anomaly, the agent can provide a summary of the impact of the anomalous behavior and next steps to take.
• Cluster and segment prospective customers: An agent can review metadata for prospective customers and identify meaningful clusters. For each cluster, the agent can summarize why each segment is different, then suggest specific actions to take for each group. For example, accelerate prospective customer outreach by prompting the agent to draft an email message and send it to a list of relevant recipients in each cluster, then call an API to update an internal outreach tracker.
• Sophisticated analysis of complex datasets: An agent can perform sophisticated analysis such as "Identify unused servers by analyzing the currently running applications and server inventory" by reasoning across multiple related data sources, even if those sources are in different connections.
• Perform scenario analysis for strategic planning: An agent can analyze possible scenarios that could affect the product supply chain by generating what-if scenarios based on user inputs and data, running scenarios using external models defined in Python code or stored procedures. When performing scenario analysis, the agent can state assumptions and uncertainty, and evaluate the output of a possible scenario compared with real outcomes provided in an input table.

Limitations of Sigma agents

When building AI agents in Sigma, consider the following limitations:

• You can trigger an agent to run as part of an automated action sequence, but not from an action sequence triggered on select or on click.
• An agent can exist in only one workbook.
• An agent does not have context about data elements that do not exist in the workbook and are not made available to the agent.

Requirements

To build a Sigma agent in a workbook, you must have the following:

• The Manage agents and Create, edit, and publish workbooks feature permissions enabled on your account type
• Can edit access to the workbook

📘

The Manage agents feature permission is enabled by default for users assigned the default Build or Pro account types, and custom account types with Can write SQL enabled.

Your Sigma organization must have an AI provider configured. See Set up an AI provider. The configured AI provider is used for all AI functionality in Sigma. Agents that you build can run with an external AI provider or a warehouse-hosted model as an AI provider.

Build an agent in a workbook

To build an agent in a workbook:

1. Open an existing workbook for editing, or create a workbook from Sigma Home by choosing Create new > Workbook.

2. In the editor panel, select Agents, then click Add agent (+).

   A Configure agent modal appears.

3. Next to the placeholder agent name (Agent 1) select the down arrow, then choose Rename to provide a descriptive name for the agent. Press Enter to save the name.

4. In the Instructions section, provide instructions and context to the agent. See Add instructions for the agent.

5. In the side panel, add one or more data elements from the workbook that the agent can use as context. See Add data for the agent.

6. In the side panel, add one or more actions or action sequences that the agent can run in the workbook. See Set up actions for the agent.

7. (Optional) If you have a Snowflake Cortex Agent or a Databricks Genie Space in your connected data platform, you can add the warehouse agent as a tool for the Sigma agent. In the side panel, for Warehouse agents, click Add warehouse agent (+). See Set up warehouse agents as tools for the agent.

8. (Optional) If you have an MCP tool configured in Sigma, you can add the MCP server as a tool for the Sigma agent. In the side panel, for MCP tools, click Add MCP tool (+). See Add an MCP tool to a Sigma agent.

9. After setting up the agent, click Preview to interact with the agent using the instructions and data that you configured.

10. When you are satisfied with your agent, click Save.

You can build multiple agents in a workbook.

Add instructions for the agent

Add instructions for the agent to provide relevant context and guidance to help the agent function how you want.

💡

At any time, you can use the Agent builder assistant (). Where prompted with How can I help you build your agent?, enter a question or prompt. See Add instructions for the agent.

When you write instructions, you can do any of the following:

• Add organization-specific context, such as the fiscal year quarters that your organization uses to conduct business or the names of subsidiary organizations.

• Define specific user context by adding dynamic text formulas that reference, for example, the current user's full name:

  When possible, address the user chatting with you by name: =CurrentUserFullName()

• Provide factual information by adding dynamic text formulas that reference the current day or time:

  If asked for the time, or when using temporal language, use the time: =Now()

• Pass additional context by retrieving values from control elements in your workbook with a dynamic text formula.

  For example, if you have a list values control with an ID chatbot-tone with 3 manually provided options: one "short and sweet", another "professional tone", and another "term paper length", reference the control value in the instructions using a dynamic text formula:

  When responding, use the following tone: =[chatbot-tone]

  Then a user can choose the desired tone of the agent using the list values control, and the instructions can pull the user selection from the control and use the selected tone when responding to the user.

• Instruct the agent with desired response details, such as output structure to use, desired response length, or information to include:

  If asked to summarize findings, respond in no more than 5 bullet points and use action verbs. Always include a title of "Summary".

• Increase the likelihood that the agent generates responsible insights by instructing the agent to cite sources:

  When asserting a fact, include a reference to the source of the fact. If you provide a link to relevant documentation page or a webpage, make sure that the page is accessible on the web before providing it.

  Also, provide explicit instructions to avoid off-limits topics:

  Never provide investing advice or guidance, even if asked for it. Instead, politely decline.

• Mention different tools configured to run actions or retrieve data, with additional guidance for the agent about when to run different tools.

  For example, guide the agent about when to use an action tool:

  If a question requests more details about a row of data, use the View Details action tool in the response.

  Or provide explicit instructions about when to use an agent from a data platform connection:

  If a question mentions the word "sales", use the "sales insights" Genie Space.

  Or provide context about when to call an MCP tool:

  To confirm that a recommendation complies with the company sales & promotion policy, use the Sharepoint MCP server "get document contents" tool to retrieve the sales and promotion policy details from the **Promotion Regulations** Microsoft SharePoint folder.

When instructions are changed, the chat refreshes.

Add data for the agent

Add data to the agent that it can use as context when answering questions and responding in chat conversations. Data sources must already be added to the workbook.

1. When creating or editing an agent, for Data elements, click Add data element (+) to add one or more data elements from the workbook to provide context to the agent. You can add a table, pivot table, chart, input table, or linked input table element in the workbook.

   When adding a data source, consider the following:

   • Data sources can be from different connections. The agent can use any of the specified data sources to respond to questions and prompts in a chat.
   • If the data source for the agent is sourced from a data model, the agent cannot access related columns or metrics available through relationships.

2. After you add data sources, revise the instructions, define action tools, click Preview to preview the changes, or click Save to save the agent and return to editing the workbook.

Set up actions for the agent

The agent can run actions in the workbook as tools. When you create an action tool for an agent, you can define a set of deterministic steps to be performed when the tool is called. Each step can do one of the following:

• Run an action.
• Trigger an action sequence defined in the workbook, if the action is associated with an "On click" action. For example, actions triggered when clicking a button.

💡

In your workbook, rename action sequences to more descriptive names to make it easier to locate specific sequences when adding them as tools for the agent.

To add an action tool for an agent:

1. When creating or editing an agent, in the side panel for Actions, click Add action (+).

2. For Name, enter a name for the tool. Specify a name that makes it clear what the tool does.

3. (Optional) For Description, enter a description of what the tool is used for with guidance about when to use the tool. For example, describe a tool sends a notification to a Slack channel about the current state of the agent conversation with: "Summarize current state of the chat conversation for teammates in Slack".

   💡

   The agent uses the name and description to decide when to use the tool.

4. (Optional) To prompt the user to approve the action in the chat conversation with the agent, turn on the Requires user approval toggle. When a user is prompted for approval, the name of the action tool is shown in the chat conversation.

   💡

   An agent can run an action tool with or without approval. The best option depends on the importance of data governance to your use case and how much autonomy you want to give the agent:

   • Restrict agent autonomy with strict human-in-the-loop review: Require user approval for the action.
   • Guided agent autonomy without review: Do not require user approval for the action. Guide the agent behavior with action tool descriptions and guidance for agent input placeholders.

5. For Steps, click Add step (+).

6. For the step, select a Step type. Choose between Run an action and Trigger an action sequence, if one exists.

   • If you choose Run an action, for Action, select the action to perform in this step, then complete the relevant fields for the action.
   • If you choose Trigger an action sequence, for Sequence, select the action sequence in the workbook that you want to trigger. The sequence must be one triggered "On click".

   💡

   If your action tool outputs data, such as by inserting or updating a row in an input table, setting form field values, or setting control values, you can choose the Agent input option to allow the agent to provide the input for the action. After selecting Agent input, you can replace the placeholder text with further instructions to the agent about what information to enter.

   For examples, see Example: Add conversational memory as an action tool

7. Repeat steps 4–6 to add more steps to the action tool.

8. After you finish adding action tools, revise the instructions. For example:

   Text

   When prompted to send an email update, use the "Notify leadership of recent updates" action tool.

9. Click Preview to preview the changes to the agent, or click Save to save the agent and return to editing the workbook.

Set up warehouse agents as tools for the agent

If you have a Snowflake Cortex Agent or a Databricks AI/BI Genie space set up in a connected data platform, your Sigma agent can call one or more to perform tasks. Using a warehouse agent like a Cortex Agent or Genie space, you can give the Sigma agent context about data that exists only in your data warehouse, and doesn't exist in Sigma or in the workbook that the Sigma agent is operating in.

💡

A Sigma agent can function as an orchestrator, instructing a set of warehouse agents to perform narrowly scoped tasks on specific data sources, and summarizing the responses to inform further analysis or take action in Sigma.

A Sigma agent respects role-based access control and can only use warehouse agents that you have access to. If the Sigma agent has access to a warehouse agent that a user chatting with the Sigma agent cannot access with their role in Snowflake or Databricks, the Sigma agent cannot use that warehouse agent.

1. When creating or editing an agent, for Warehouse agents, click Add warehouse agent (+).

2. Search for and select one or more warehouse agents.

   📘

   If you do not see any warehouse agents, or an agent that you expect to see is not listed, make sure that your connection is updated and that you are using a role with access to the agent. If you sync the connection, you must sync both the database or catalog and schema that contain the agent. For more details, see Manually sync your data.

3. After you add warehouse agents, revise the instructions. For example, add instructions for when to use a warehouse agent:

   Text

   When prompted for details about recent sales, use the RECENT_SALES agent.

4. Click Preview to preview the changes to the agent, or click Save to save the agent and return to editing the workbook.

Add an MCP tool to a Sigma agent

To add an MCP tool to a Sigma agent, do the following:

📘

To add an MCP tool to a Sigma agent, you must be assigned an account type with the Use MCP tools permission enabled and you must have Can use access to the MCP tool.

1. Open a workbook for editing.

2. In the editor panel, select Agents.

3. For the relevant agent, select More > Edit.

4. For MCP tools, click Add MCP tool (+).

5. Search for and select the relevant MCP tool.

6. Update the instructions for the agent with guidance for using the MCP tool.

7. (Optional) Select Preview and enter a prompt to test the agent's ability to use the new MCP tool.

8. After you add MCP tools, revise the instructions. For example:

   Text

   When asked questions about the latest brand styles, use the Microsoft 365 MCP tool to connect to SharePoint and retrieve details from the Branding folder.

9. Click Preview to preview the changes to the agent, or click Save to save the agent and return to editing the workbook.

For more details about MCP tools, see Configure MCP tools.

Add a chat element to interact with an agent

After building an agent, add a chat element to allow users to interact with the agent:

📘

To add a chat element, you must have Can edit access to the workbook and the Create, edit, and publish workbooks feature permission enabled on your account type. You do not need the Manage agents permission to add a chat element.

1. Open a workbook for editing.

2. In the Add element bar, select UI > Chat.

3. Select the chat element.

4. In the Properties tab of the editor panel, select an agent.

5. Publish the workbook to save the default agent selection.

   The chat element displays content relevant to the instructions, inviting a user to interact with the chat element. See Chat with an agent in Sigma.

Call an agent on a schedule with an action

You can call an agent as part of an automated action sequence to automate multi-step workflows that gather data, generate structured output, and trigger downstream actions without any user interaction.

When an automated action sequence calls an agent, the agent runs once with the prompt included with the configured action, then runs its available tools (such as querying data sources), and returns structured output. The output is structured as action variables that you can reference in following actions in the sequence.

A typical automated action sequence that calls an agent might look like the following:

1. Call agent: The agent receives instructions to query multiple data sources and return structured output, such as a summary and description.
2. Update input table: An insert or update row action writes the agent's output—passed as action variables—to an input table for a persistent record.
3. Send notification: A notification action uses the action variables to send the summary and description to a Slack channel, Microsoft Teams, or email.

Add a Call agent action to an automated action sequence

To call an agent on a schedule, set up an automated action sequence with the Call agent action:

1. Follow the steps to Run action sequences automatically.

2. In the action configuration popover, select the Call agent action:

   1. For Agent, select an agent in the workbook.
   2. For Prompt, enter a prompt to send to the agent when the action runs. Include enough details to make sure that the agent can produce desired results without follow-up questions. You can test your prompt in the chat element.
   3. For Output, define action variables to structure the output from the agent. By default, the Agent-response action variable is defined to make the entire response of the agent available as text. Click + Add variable to further structure the output. The agent uses instructions in the prompt and the name of the action variables when structuring output.

3. After configuring the action, finish setting up the automated action sequence.

Example: Nightly usage summary agent

In this example, an agent runs every night to review the last 24 hours of usage data across multiple sources and send an email with the summary.

The agent is configured with the following:

• Multiple sources of usage data in the workbook for data from the last 24 hours.
• Instructions to provide a concise summary for a business systems analyst.

To set up this workflow:

1. Create an automated action sequence with the following actions:

   1. Call agent action with the following configuration:

      1. Select the agent.

      2. For Prompt, enter guidance to provide a concise summary of the latest usage data to send via email, with an action-oriented email subject. For example, Provide a 3 sentence summary of usage data patterns for a business systems analyst and write an email subject that uses a call to action.

      3. For Output, add the following action variables:

         • email-subject with a data type of Text.
         • email-body with a data type of Text.

   2. Notify and export action with the following configuration:

      1. Select a destination of Email, then choose recipients.
      2. For Subject, press =, then type [email-subject] to reference the email subject action variable from the Call agent action.
      3. For Message, press =, then type [email-body] to reference the email body action variable from the Call agent action.
      4. Complete the remaining configuration options.

2. After configuring the action sequence, follow the steps to run the action sequence automatically at your preferred frequency, such as every morning at 6AM.

3. Publish the workbook to activate the schedule.

Example agent implementations

You can use Sigma agents to perform actions and help you accomplish tasks. Refer to the following example implementations:

• Example: Add conversational memory as an action tool
• Example: Notify a user when an anomaly is detected
• Example: Run a forecasting model
• Example: Retrieve details from a third-party API

Example: Add conversational memory as an action tool

If you want to help the agent remember important context from conversations, add an input table to your workbook and create an action tool for the agent to use.

1. Add an input table to the workbook titled Conversation memory. Add one date column called Date, and two text columns called Conversation summary and Key detail.

2. Select an agent and follow the steps to add an action tool:

   1. For Name, enter Remember details.

   2. For Description, enter Keep track of important details in a conversation when prompted, adding a short summary and the key detail to remember.

   3. For Steps, click Add step (+).

   4. For the step, select a Step type of Run an action.

   5. Choose the Insert row action and select the Conversation memory input table.

   6. For Map with values, do the following:

      • For the Date column, select Formula and enter Now() to record the date when the key detail was added to the table.
      • For the Conversation summary column, select Agent input. Replace the placeholder text with Conversation summary.
      • For the Key detail column, select Agent input and replace the placeholder text with Key detail to remember.

   7. Select Close (x) to return to editing the agent.

3. Update the instructions with guidance for using the new action tool. Add a sentence like When asked to remember a key detail, use the remember details action tool.

4. (Optional) If you want the agent to reference the key details stored in the Conversation memory input table, add the input table as a data element, then click Settings () and turn off the Allow write access toggle. By adding the table and turning off write access, the agent can read the contents of the table and use it as context, but only write to the table using the action tool.

5. Click Save to save the agent.

6. Add a chat element to interact with the agent.

💡

Using this method creates conversation memory that is shared with all users of the workbook. If you want the agent to store key details specific to each user, add a Created by system column from the Row edit history, then follow the steps to set up row-level security using the user email address.

Example: Notify a user when an anomaly is detected

For example, if you want to monitor anomalous page visits to your website, you might have a workbook with your website analytics data.

You can configure an alert that sends a Slack message when page views exceed a known threshold, but you can also use a Sigma agent to review non-deterministic anomalies and send an alert.

You can recreate this example with the Sigma Sample Database Google Analytics Events data.

1. In a workbook, add the Sigma Sample Database Google Analytics Events table and a text area control titled Message body.

2. Create a Sigma agent.

3. Add a data element with the website analytics data table. If you use complex calculations to identify page views or unique users, use a table with metrics for those calculations defined.

4. Provide instructions to the agent with guidance like the following:

   Text

   You are a website analytics power agent. Review the provided data sources and identify anomalous activity, such as a spike in the number of page views compared to previous days or previous hours, an unexpected increase in unique users early in the morning for a Pacific time zone, or a much lower number of engaged sessions than usual.
   
   Refer to the company marketing calendar stored in Microsoft SharePoint for details about launch announcements or social media activity to help explain anomalies.
   
   After identifying an anomaly, do the following:
   
   1. Summarize details about the anomaly and write an explanation about why it occurred.
   2. Generate simple HTML to style the details and explanation as a message, then use the **Stage message contents** action. Never use placeholder values.
   3. When the message is ready to send, call the **Notify about anomalies** action tool to notify a human about the anomalous traffic.

5. To store the message contents, follow the steps to add an action tool to update the text area control:

   1. For Name, enter Stage message contents.
   2. For Description, enter Write simple valid HTML to provide as a Microsoft Teams message. Refer to https://learn.microsoft.com/en-us/microsoftteams/platform/bots/how-to/format-your-bot-messages for supported syntax.
   3. For Steps, click Add step (+).
   4. For the step, select a Step type of Run an action.
   5. Choose the Set control value action.
   6. For Update control, choose the Message body text area control.
   7. For Set value as, select Agent input. Optionally update the placeholder text and enter Message body.
   8. Select Close (x) to return to editing the agent.

6. To send a notification about identified anomalies, follow the steps to add an action tool:

   1. For Name, enter Notify about anomalies.
   2. For Description, enter When you identify an anomaly, notify a human in Microsoft Teams.
   3. For Steps, click Add step (+).
   4. For the step, select a Step type of Run an action.
   5. Choose the Notify and export action and select a Destination of Microsoft Teams.
   6. For To, choose Specific channels and enter the URL to the relevant alert Microsoft Teams channel.
   7. For Message, enter =[message-body] to reference the text area control updated by the agent.
   8. (Optional) Turn off the Link to workbook toggle.
   9. (Optional) Turn off the Attachment toggle.
   10. Select Close (x) to return to editing the agent.

7. Click Save to save the agent.

8. Add a chat element to interact with the agent, or schedule the agent to run to notify a user on a regular cadence.

Example: Run a forecasting model

If you want to combine deterministic forecasting models with agent-informed forecasting, you can add a Sigma agent to a workbook with an existing Python code element that you use to run a forecasting model.

1. In a workbook, add a Python element with the code for the forecasting model. Include the sigma.output() method to make the code available as a child element, and create a table with the code output. Add another table with relevant inputs for the forecasting model, such as a historical product inventory table.

2. Create a Sigma agent.

3. Provide instructions to the agent with guidance like the following:

   Text

   You are an expert at forecasting inventory. If you need to retrieve deterministic inventory forecasts, run the inventory forecasting model and review the results. The output is available in the Inventory Output table.

4. Add both the historical product inventory table and the table containing the output as a Data element for the agent.

5. To make the agent capable of running the forecasting model, add an action tool to run the Python element:

   1. For Name, enter Run forecasting model.
   2. For Description, enter Retrieve deterministic inventory forecasting using a trend projection technique.
   3. For Steps, click Add step (+).
   4. For the step, select a Step type of Run an action.
   5. Choose the Run Python element action.
   6. For Element, choose the Python element that contains the inventory forecasting model.
   7. Select Close (x) to return to editing the agent.

6. Click Save to save the agent.

7. Add a chat element to interact with the agent.

Example: Retrieve details from a third-party API

A Sigma agent can retrieve details from a third-party API, summarize the response, and use the response to take further action in Sigma.

For example, if you want to build a Sigma agent to help search for an apartment in New York City, NY, USA, you might want to inform your search with data from 311 to identify whether and what type of incidents are commonly reported for the apartment building address.

1. In your Sigma organization, configure API credentials and connectors for the NYC Open Data API.

2. In your workbook, add a text input to use for entering relevant apartment addresses, an input table to store recommendations from the Sigma agent.

3. Create a Sigma agent.

4. Provide instructions to the agent with guidance like the following:

   Help me look for an apartment! Review information available to you on the web and also in 311 to make recommendations about specific apartment building addresses. If I ask you to look at 311 data, use the "311 incidents" action tool.

5. For Data elements, add the Recommendations input table and leave the Allow write access toggle turned on.

6. To make the agent capable of calling an API with 311 incident data in NYC, add an action tool to call an API action:

   1. For Name, enter Retrieve 311 incidents.

   2. For Description, enter Call 311 with the specified address and summarize the response.

   3. For Steps, click Add step (+).

   4. For the step, select a Step type of Run an action.

   5. Choose the Call API action.

   6. For Select an API connector, choose the NYC Open Data connector.

   7. For Map with values, if any API parameters are dynamically set, such as the incident address, choose Control and specify the text input control.

      💡

      To let the agent provide the input, select Agent input instead of Control.

   8. Select Close (x) to return to editing the agent.

7. Click Save to save the agent.

8. Add a chat element to interact with the agent.

Chat with an agent in Sigma

If you have access to a workbook that contains a chat element, you can chat with an agent to ask questions of the data in the workbook, accomplish tasks, enrich your understanding with context from third-party systems, and more.

User requirements

You can chat with a Sigma agent if you have at least Can view access to a workbook that contains a chat element and the View workbooks permission enabled on your account type.

You can chat with an agent, and choose which agent you chat with (if multiple exist) if you have at least Can explore access to a workbook and the Full explore permission enabled on your account type.

Chat with an agent in a workbook

To chat with a Sigma agent in a workbook:

1. Open a workbook with a chat element.

2. When the workbook page with the chat element loads, the agent describes what it can do and what it has access to.

3. Enter a prompt or query to chat with the agent.

   Some actions that the agent can perform might prompt you for approval. To allow the agent to perform the action, click Approve, or to skip performing the action, click Skip.

   If the agent uses an MCP tool that requires authorization, such as one that accesses Atlassian services, you are prompted to Authorize the MCP tool so the agent can perform actions with your access.

If you have Can explore access to the workbook, you can change which agent you can chat with:

1. In the workbook, click Customize this view.
2. Select the chat element.
3. In the Properties tab of the editor panel, for Agent, select the dropdown menu to choose a different agent.
4. Return to the chat element and start chatting with the new agent. When you exit the custom view, the agent selection returns to the agent available in the published version of the workbook.