How to Set Up Sigma Agents for Data Workflows

Sigma Agents are AI-powered workflow automations built into Sigma Computing that connect directly to your cloud data warehouse and take action on live data. Setting one up requires an existing connection to Snowflake, Databricks, or a major cloud warehouse, admin-level access inside Sigma, and a clear definition of what data to monitor and what action to trigger. Conversational agents went live on April 2, 2026. Autonomous agents are rolling out through the year.

What Sigma Agents Do

Sigma Computing launched Sigma Agents on April 2, 2026 as part of its Spring Product Launch, following a preview at the Workflow 2026 user conference in San Francisco. The core premise differs from traditional BI: instead of waiting for someone to open a dashboard and notice a problem, an agent watches your live warehouse data continuously and acts when a defined condition is met.

A practical example: an operations team connects an agent to a Snowflake table tracking inventory across a retail chain. When stock on a high-velocity SKU drops below a reorder threshold, the agent fires a Slack alert to the procurement team and opens a Jira ticket with the item details, without anyone running a query or refreshing a report.

According to Sigma, agents can write data back to the warehouse, trigger updates in CRMs and ticketing systems, send alerts, and call external APIs via webhooks. Security is inherited directly from the connected warehouse. Existing row-level security and permission structures apply automatically, which removes the need to configure a separate access layer for the agent.

Sigma Computing doubled its ARR in the 12 months leading up to its first user conference in March 2026, a period the company attributed to market demand shifting from passive dashboards toward AI-powered workflows that act on live data.

The Three Agent Modes

Sigma organizes agents into three modes, each suited to a different level of human involvement.

Conversational agents are available now. A user opens a natural-language interface inside Sigma and queries an agent directly. The agent retrieves data from the connected warehouse, reasons through the question with visible chain-of-thought logic, and presents a result. Before executing any write operation or external action, the agent requests human approval. This makes it well-suited to tasks where oversight is required during an early rollout.

Human-in-the-loop agents are built for workflows that require manual sign-off before execution. An expense management agent might detect an anomalous charge against a software budget, prepare a summary, and pause before sending an alert or updating a record, waiting for a finance manager to review and confirm. This mode handles high-stakes decisions where automated action without review carries real operational risk.

Autonomous agents run on a defined schedule without human input. They scan live data on a timer, apply threshold or anomaly logic, and execute configured outputs when conditions are met. Autonomous agents are rolling out through 2026. Once available, an agent checking overnight sales figures at 6 a.m. could identify regional underperformance and send a briefing to the relevant lead before the workday begins, with no human initiation required.

What You Need Before You Start

Three prerequisites apply to every Sigma Agent.

A connected cloud data warehouse. Sigma Agents work with Snowflake, Databricks, and cloud warehouses on AWS, Azure, and Google Cloud via PrivateLink or Private Service Connect. The agent reads from this connection and, depending on configuration, writes back to it. There is no file-import path for agents. The warehouse must be live and already connected to a Sigma workspace before building an agent.

Admin-level or delegated permissions. Because agents can execute write operations and call external APIs, building one requires appropriate permissions inside the Sigma admin console. Existing warehouse roles pass through Sigma's inherited security model and apply to the agent automatically, so no separate permission layer needs to be configured.

A well-defined problem. Agents perform best when built around a specific monitoring condition, a clear trigger, and a configured output. Teams that define all three before opening the agent builder configure working agents faster than those who start without a clear goal. A monitoring condition like "weekly software spend exceeds budget by 10 percent" is immediately actionable. A condition like "watch spending" is not.

If your data lives in spreadsheets or local exports rather than a cloud warehouse, VSLZ AI offers a way to build similar monitoring and analysis workflows from a file upload with no warehouse connection or SQL required.

How Sigma Agents Are Structured

Every Sigma Agent is built on four components: Instructions, Data, Tools, and Outputs.

Instructions are natural-language text describing what the agent monitors, what logic it applies, and how it communicates results. Specific instructions produce more reliable behavior. "Monitor daily revenue by region and alert if any region falls more than 15 percent below its 30-day rolling average" will produce consistent results. "Watch sales" will not, because the scope is undefined.

Data defines which warehouse tables, workbooks, or datasets the agent can access. Scoping the data surface to only what the agent needs improves consistency and reduces the chance of the agent acting on irrelevant signals. An inventory agent should be pointed at the inventory table, not the entire schema.

Tools are the actions the agent can take: writing to the warehouse, querying connected data, calling external APIs, and sending notifications via Slack. The 2026 roadmap includes expanded API connectors and support for MCP (Model Context Protocol) servers, which will allow agents to interact with a broader range of external systems. An external API for calling Sigma agents from outside the platform is on the roadmap for later in the year.

Outputs define what the agent produces: a written summary, a data update, a ticket in a connected tool, or a message sent to a channel. Outputs can require human approval before executing, which is the mechanism that distinguishes human-in-the-loop behavior from fully autonomous operation.

Use Case Examples

Financial planning and spend monitoring. An agent watches budget versus actual spending across departments, sends weekly variance summaries to finance leads, and escalates when a category exceeds its allocation by a defined threshold. Finance teams using this pattern replace a recurring manual process of downloading reports and comparing against plan figures.

Retail and inventory operations. An agent monitors stock levels across SKUs, triggers purchase orders or Slack notifications when inventory drops below reorder points, and logs the action in a tracking table in the warehouse. This covers the gap between a warehouse data update and the procurement team becoming aware of the change.

Customer support escalation. An agent monitors a support ticket table for cases meeting escalation criteria, such as tickets open longer than 48 hours or those tagged as critical, and creates follow-up tasks in Jira automatically. Support leads stop checking dashboards for overdue tickets and receive direct, condition-triggered notifications instead.

Sales pipeline hygiene. An agent scans a pipeline table nightly, identifies deals that have gone more than a defined number of days without activity, and sends a prompt to the assigned account owner via Slack or email. Teams using this pattern report fewer deals falling through the cracks during periods of high volume.

What Is Available and What Is Coming

Conversational agents and human-in-the-loop agents are available as of April 2026. Autonomous scheduled agents are rolling out through the year. The Sigma MCP Server, which will allow external AI systems to interact with Sigma workbooks and agents programmatically, is on the roadmap for later in 2026.

Teams can start with conversational agents today and build out their Instructions, Data, and Tools configuration in anticipation of autonomous scheduling becoming available. The four-component structure is consistent across all agent types, so work done configuring a conversational agent transfers directly when autonomous agents launch. Starting with a well-scoped monitoring problem, a clear trigger, and a defined output will produce results regardless of which agent mode is in use.