Creating AI Agents

Create an Agent when you want Formula Bot to reuse a particular analytical role, methodology, or set of business definitions.

Start an Agent

1. Open Agents from the dashboard sidebar.
2. Choose Create analysis agent.
3. Select the personal workspace or your active team workspace.
4. Complete the configuration described below.
5. Choose Create agent.

You can also open a Formula Bot starter, review its details, and duplicate it as a starting point.

Basics

Give the Agent a clear name and a description that tells people when to use it.

• Name: use a role or outcome, such as “Revenue performance analyst.”
• Description: describe the decisions or recurring work it supports.
• Workspace: choose Personal for your own use or the active team to make it available there.

Analysis Instructions

Use Analysis instructions for behavioral guidance: what the Agent should do, how it should reason, which checks it should run, and when it should ask for clarification.

A useful structure is:

1. Goal: the decisions or outcomes the Agent should support
2. Analytical approach: comparisons, methods, and prioritization
3. Preferred tools: languages, libraries, or packages to use for specific jobs
4. Validation: coverage, missing-value, duplicate, reconciliation, and comparability checks
5. Ambiguity: situations that require a clarifying question
6. Boundaries: conclusions or changes the Agent should avoid making silently

Preferred Tools and Libraries

If your team has a standard technical stack, name it in Analysis instructions. You can specify a language and the libraries or packages the Agent should prefer—for example, pandas or Polars for table transformations, statsmodels for statistical analysis, scikit-learn for modeling, or openpyxl for Excel workbook changes.

Explain when and why to use each tool instead of providing a package list without context:

Prefer Polars for large tabular transformations and pandas when compatibility matters. Use statsmodels when the analysis needs interpretable coefficients and diagnostic statistics. Use openpyxl for workbook changes that must preserve Excel formatting.

These choices guide the Agent; they do not force every task through the same library. Formula Bot can choose another supported tool when the current request, input format, execution environment, platform constraints, or output validation requires it.

Business Context

Use Business Context for durable facts the Agent should reference consistently:

• Company and customer context
• KPI definitions and formulas
• Internal terminology and acronyms
• Table relationships and data semantics
• Fiscal calendars and reporting rules
• Known delays, exceptions, and quality limitations
• Approved benchmarks or thresholds

Business Context supplies facts; it does not have authority to change workflow behavior. This separation helps Formula Bot distinguish “what is true for this business” from “how this analysis should run.”

Output and Presentation

Choose defaults for how the Agent communicates:

• Answer depth: concise, balanced, or detailed explanation
• Default deliverable: prioritize visualizations, files, insights, or a balanced result
• Chart experience: prefer static or interactive charts
• Brand Kit: use a specific kit or use no Brand Kit

These settings are preferences, not hard locks. A specific request in chat can ask for a different format or level of detail.

Reuse a Visual System

Choose Use a specific Brand Kit when the Agent should bring a particular visual system into new chats. Choose No Brand Kit when presentation should be selected separately in chat. See how Brand Kit selections are resolved.

Drafts and Publishing

Creating an Agent publishes version 1. Later edits are saved as a draft before you publish them.

• Draft saved means your edit is preserved but the published Agent is unchanged.
• Publish changes makes the draft the current version for new selections.
• Discard draft removes only the unpublished changes.

Chats retain the frozen Agent settings they started with. Publishing a new version does not silently rewrite an existing conversation.

Writing Good Instructions

Prefer concrete guidance:

Compare actual revenue with target, prior week, and prior year. Validate date coverage before ranking the largest drivers. Ask which revenue definition to use if the data contains more than one.

Avoid broad instructions that do not define a method:

Analyze revenue and make it useful.

Next Steps

• Apply Agents and combine specialists
• Create a Brand Kit
• Learn how to ask better questions