What happens when an AI agent edits a spreadsheet? Inside headless LibreOffice's silent OOXML rewrites

The wrong question

Evaluating an AI agent on how well it formulates formulas or generates Python code misses the critical failure boundary. In production agent workflows, the primary point of failure is the file runtime.

When an agent needs to read, recalculate, or update a workbook, developers face three flawed compromises:

To edit a single cell, containerized agent harnesses resort to the first option: headless LibreOffice.¹ Here is what happens next:

Silent package drift: diffing the raw OOXML

An Excel workbook (.xlsx or .xlsm) is an Open Packaging Conventions (OPC) zip archive containing dozens of interconnected XML parts: worksheet data, shared string tables, styles, themes, drawing relationships, chart definitions, and pivot caches.

Now consider this: when an agent updates cell B2, only cell B2 changes. Styles tables, drawing relationship trees, and theme palettes remain untouched: a surgical patch directly into the ZIP archive.

We tested what happens when an agent modifies a single cell across five real-world spreadsheet fixtures:

• A financial reporting model with chart linkages and drawing parts.
• A template with external hyperlinks and custom document properties.
• A formatted workbook with custom theme palettes and conditional styling.
• A high-density table with shared string references.
• A macro-enabled operational pack (XLSM).

Here is the same workbook, saved twice. The only requested change in both runs was cell B2, which lives in xl/worksheets/sheet1.xml:

Representative part set from the preservation receipts. Amber M = the part containing the edited cell. Rose M = an untouched part rewritten anyway. Dotted = byte-identical. Across the five-fixture study the unrelated-part count ranged from 8 to 151.

The observation: On a single-cell save, headless LibreOffice rewrote between 8 and 151 untouched internal parts across the fixtures.² It re-serialized XML trees, dropped non-standard namespaces, reordered style index tables, and rebuilt drawing relationship graphs.

@@ -1,12 +1,10 @@
- <xdr:wsDr xmlns:xdr="http://schemas.openxmlformats.org/drawingml/2006/spreadsheetDrawing"
-            xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main"
-            xmlns:r="http://schemas.openxmlformats.org/officeDocument/2006/relationships">
+ <xdr:wsDr xmlns:xdr="http://schemas.openxmlformats.org/drawingml/2006/spreadsheetDrawing" xmlns:a="http://schemas.openxmlformats.org/drawingml/2006/main">
    <xdr:twoCellAnchor editAs="oneCell">
-     <xdr:from><xdr:col>4</xdr:col><xdr:colOff>0</xdr:colOff><xdr:row>2</xdr:row><xdr:rowOff>0</xdr:rowOff></xdr:from>
-     <xdr:to><xdr:col>9</xdr:col><xdr:colOff>12700</xdr:colOff><xdr:row>16</xdr:row><xdr:rowOff>50800</xdr:rowOff></xdr:to>
-     <xdr:graphicFrame macro="">
-       <xdr:nvGraphicFramePr><xdr:cNvPr id="2" name="Chart 1"/><xdr:cNvGraphicFramePr/></xdr:nvGraphicFramePr>
-       <a:graphic><a:graphicData uri=".../chart"><c:chart xmlns:c="..." r:id="rId2"/></a:graphicData></a:graphic>
+     <xdr:from><xdr:col>4</xdr:col><xdr:row>2</xdr:row></xdr:from>
+     <xdr:to><xdr:col>9</xdr:col><xdr:row>16</xdr:row></xdr:to>
+     <xdr:graphicFrame>
+       <xdr:nvGraphicFramePr><xdr:cNvPr id="1" name="Object 1"/></xdr:nvGraphicFramePr>
+       <a:graphic><a:graphicData uri=".../chart"><c:chart xmlns:r="..." r:id="rId1"/></a:graphicData></a:graphic>
    </xdr:twoCellAnchor>

@@ -14,5 +14,5 @@
     <row r="2" spans="1:4">
       <c r="A2" t="s"><v>0</v></c>
-      <c r="B2"><v>1000000</v></c>
+      <c r="B2"><v>1250000</v></c>
       <c r="C2" s="1"><f>B2*0.2</f><v>200000</v></c>
     </row>
# 0 other files modified (styles, drawings, themes, and relations unchanged)

Latency on real workloads: 12-0 benchmark sweep

Spawning a 410 MB desktop suite inside an agent container also carries substantial execution latency. We benchmarked WolfXL 2.1.0 against LibreOffice 24.2.7.2 on a dedicated 16-vCPU x86_64 Linux system using a 200,000 row by 8 column dataset across 12 registered task routes.³

Across every tested route, WolfXL outputs matched expected values.⁴ Cold start, warm latency, bulk CSV, roundtrip save, and PDF conversion were all won by WolfXL in this measured matrix.

Container footprint and process overhead

For autonomous agent infrastructure, disk footprint and process initialization directly govern container provisioning speed, ephemeral disk costs, and memory ceiling under concurrency.

Reproducibility and evidence receipts

All benchmark scripts, synthetic fixtures, diff tools, and SHA-256 evidence receipts are public in the WolfXL open-source repository so you can inspect the exact methods and verify numbers on your own hardware.