Skip to main content
Agent Mode (shown in the app as the Agentic Loop) is a mode of the AI Assistant that turns a single request into an autonomous, multi-step run: the AI plans a goal, executes one step, reads the result, and decides the next step on its own — repeating until the goal is done.
This page documents Agent Mode specifically. For the AI Analyst sidebar in general (Chat mode, voice input, quick insights), see AI Assistant.

​
Turning On Agent Mode

In the AI Analyst sidebar, switch the mode selector from Chat to Agent before submitting your prompt. Once submitted:
  • The badge Agentic Loop Active appears
  • The sidebar shows a live status such as “Cleaning data… (Step 2)” or “Generating chart… (Step 3)”
  • You can stop the loop at any time with the stop button

​
How the Loop Works

1

You give a goal

e.g. “Clean this data, calculate average profit by region, and chart it.”
2

Step 1 runs

The AI writes and runs code for only the first part of the goal (e.g. cleaning).
3

The AI reads its own result

The output (or error) of Step 1 is fed back to the AI automatically with the instruction: “Continue to the next step. Fix any errors, or output [TASK_COMPLETE] if done.”
4

Step 2, 3... run automatically

The AI keeps going step by step — one code block per step — using each result to decide what’s next.
5

The loop ends

Either the AI outputs [TASK_COMPLETE] (goal finished), or the loop hits its step limit.
The Agentic Loop is capped at 5 steps. If the goal isn’t finished by then, the loop stops automatically and the sidebar shows: “Agentic loop reached maximum steps limit.”

​
Self-Correction

If a step’s code throws an error, that error is included in the next step’s context, and the AI’s next step is specifically to fix the error rather than move forward — so a single bad step doesn’t necessarily end the run.

​
Everything Agent Mode Can Do

In Agent Mode, the AI always has access to every capability category at once (in Chat mode, only the category matching your specific wording is loaded, to save time) — so it can freely combine any of the following across its steps:
Using update_cells() to edit in place (never rebuilding the table):
  • Remove duplicate rows (keeping the first occurrence)
  • Trim, uppercase/lowercase, or standardize text
  • Strip footnote brackets (e.g. [1], [a]), encoding garbage, and other unwanted symbols — while preserving things like year ranges (2023-2024)
  • Fix blank/missing values
  • Delete specific rows ({'action': 'delete_row', 'row': N})
Using insert_col() / insert_row() to modify structure in place:
  • Insert a new calculated column next to an “anchor” column (e.g. a Profit Margin column next to Revenue/Cost)
  • Append a new total/summary row
  • Move a column (insert at the new position, then blank out the original)
  • Split a single messy column into multiple columns, detecting delimiters (commas, pipes, underscores, multiple spaces, etc.) automatically
Using sort_table() / apply_filter() / clear_filter() on the existing table in place:
  • Sort ascending or descending by any column
  • Apply one or more filters (equals, not-equals, contains, greater/less than, empty/not empty, starts/ends with, or matches a list of values)
  • Combine multiple filters (e.g. Status = Active and Price > 100)
  • Clear one filter or all filters
Using apply_format() — changes appearance only, never the underlying values:
  • Background/text color (e.g. highlight negative numbers in red)
  • Number formats: currency (multiple symbols), percentages, dates
  • Bold, italic, underline/strikethrough, font size, font family, alignment, borders, decimal places
  • Can be applied to a single range or looped across the entire sheet
Using create_chart():
  • Picks an appropriate chart type from all 26 supported types (see Charts) based on your request and the data shape
  • Can chart existing columns directly, or compute new series first (e.g. extracting “Month” from a Date column) when the category you asked for doesn’t exist yet
  • Dashboards: if you ask for a “dashboard” (e.g. “create a sales dashboard”), the agent calls create_chart() 3–4 times in one step, all charts kept on-topic to the type of dashboard requested (a “profit dashboard” gets profit-focused charts, not generic ones) — see Dashboard
  • Forecasting: for “forecast” or “future trend” requests, it fits a simple linear trend (via numpy.polyfit) to historical data and charts the projected values alongside the historical ones
For analytical questions (“why did X happen”, “summarize this”, “explain the trend”), the agent does not modify the sheet. It calculates real figures in Python and prints a “Key Insights” style report: several bullet points, each a specific, numbers-backed observation (not generic advice), covering every relevant metric in the data — not just two or three cherry-picked ones.

​
Automatic Chart Suggestion After a New Table

Whenever any step (agent or chat) creates a brand-new table with set_data() — for example a groupby/pivot result — Tablix automatically evaluates whether to add a chart beside it:
  • Skipped if the table has more than 60 rows
  • Skipped if the table is only 1 row or 1 column
  • Otherwise, a chart type is picked automatically:
    • Dates in the first column → Line (or Stacked Area if multiple value columns)
    • A single category column with a small number of rows → Pie or Doughnut
    • More than 10 categories → Bar (or Stacked Bar with multiple value columns)
    • Multiple value columns otherwise → Stacked Column
    • Default → Column
  • Titles and axis labels are generated automatically from the column headers

​
Data-Handling Guardrails the Agent Follows

These aren’t things you do yourself, but they shape what you’ll see the agent produce:
  • Numeric columns are cleaned of currency symbols, %, and commas before any math is run on them, to avoid silent NaN results
  • Sums/averages are computed on full column ranges (e.g. q.cells("A:E")), not hardcoded row ranges, so results stay correct if you add more rows later
  • “Highest/top/best [metric] by [category]” is always interpreted as grouped total (SUM), not the single largest individual row, unless you specifically ask for a single transaction or an average
  • Empty rows are dropped before aggregation/charting, but preserved during column-splitting (to keep row alignment intact)
  • Every step ends with a short, plain-language summary of what changed — not a dump of the raw output data

​
Stopping and Limits

  • Click the stop button at any time to cancel the current step and end the loop; a “Stopped by user.” message is added to the chat
  • The loop has a hard ceiling of 5 steps per run — for very large multi-part goals, you may need to continue with a follow-up prompt after it stops