Ranked #1 on SpreadsheetBench: Univer Scores 68.86% in the Latest Evaluation

In the latest round of SpreadsheetBench evaluations, UniverAgent achieved a 68.86% pass rate, ranking first on the global leaderboard and outperforming mainstream solutions including ChatGPT Agent and Excel Copilot.

SpreadsheetBench

SpreadsheetBench is one of the most authoritative public benchmarks in spreadsheet automation. Its evaluation results have been referenced in official articles by both Microsoft and OpenAI, and Microsoft CEO Satya Nadella has highlighted its findings on social media.

Satya Nadella

Unlike most solutions on the leaderboard that rely on the Excel runtime environment, UniverAgent is powered by the fully self-developed Univer SDK, enabling spreadsheet computation and reasoning independent of Excel.

This represents a fundamental rethinking of what spreadsheet intelligence can be — achieved not by simply using a stronger base model, but by rebuilding the concept of a “spreadsheet agent” from the ground up:

from a tool that generates scripts to modify files, to an autonomous agent that operates in an online environment with mixed execution capabilities.

This article breaks down how UniverAgent achieves superior understanding, accuracy, and positioning performance on SpreadsheetBench, and how its system architecture departs from conventional approaches.

1. SpreadsheetBench: A Stress Test Built on Real-World Tasks

Many AI demos look impressive until they encounter real business workflows. SpreadsheetBench matters because it provides a rigorously designed benchmark grounded in tasks that genuinely occur in practice.

It includes 912 real-world tasks, spanning data cleaning, complex calculations, formatting operations, and multi-step workflows. Unlike toy datasets, SpreadsheetBench focuses on three interdependent dimensions:

1. UNDERSTANDING

Can the agent correctly interpret natural language instructions and identify the relevant regions of data — rather than acting on the wrong table or range?

2. ACCURACY

Are results numerically correct?

Are formulas logically valid?

Are types such as dates and currencies handled correctly?

3. POSITIONING

Are outputs written to the correct sheet, range, and cell — without corrupting the existing spreadsheet structure?

These three dimensions are deeply intertwined.

• A Python script may compute accurate values but fail at positioning.
• An Excel formula can position results but struggle with complex logical reasoning.
• Instruction-following may work on small examples but fall apart with large datasets.

UniverAgent performs well across all three because its system design explicitly balances them.

2. The Core Advantage: System Engineering Over Model Scale

A quick glance at the leaderboard shows a wide spectrum of technical routes. UniverAgent’s distinguishing factor is this:

It does not treat a spreadsheet as a CSV file.

It does not treat it as a generic API object.

It treats a spreadsheet as what it truly is:

a rich, structured, interactive, online computational medium.

This leads to a full-stack agent architecture built on three pillars:

1. Spreadsheet-friendly contextualization through SpreadsheetOverview and GetRangeData
2. A CodeAct + Planning execution loop instead of brittle one-shot scripts
3. An online hybrid execution engine combining Python for heavy computation and JavaScript (Univer SDK) for precise operations

3. Context Engineering: The “Map” and the “Magnifying Glass”

A major challenge in spreadsheet tasks is context explosion.

Turning a sheet with thousands of rows into plain text destroys structure, wastes token budget, and loses formulas, merged cells, style cues, and relational information.

UniverAgent uses two layers of contextual abstraction to solve this:

3.1 SpreadsheetOverview — the global “map”

SpreadsheetOverview provides a structural summary of the entire workbook at minimal token cost, using intelligent context compression.

Example output:

Key techniques include:

• Structure detection to identify table boundaries
• Hierarchical representation (Workbook → Sheet → Table → Formula)
• Formula-group aggregation (compressing 100s of formula cells into one line)

Even workbooks with over 7,000 rows usually compress to under 0.5KB of context, dramatically improving UNDERSTANDING.

3.2 GetRangeData — the structured “magnifying glass”

When deeper inspection is required, GetRangeData provides structured metadata, including:

• Values → convertible into pandas.DataFrame
• Formula groups → consistent aggregation structure
• Styles → separated semantic style layers (color, bold, number format)

This layered context design enables both global awareness and local precision.

4. Execution Strategy: From One-Shot Scripts to CodeAct Loops

Handling complex tasks in SpreadsheetBench with a single, “perfect” one-shot script is usually unrealistic. UniverAgent adopts a CodeAct (Code-as-Action) architecture, breaking execution into an iterative loop:

Thought → Code → Observation → Adjustment

In each iteration, the agent updates its internal understanding based on the latest Observation.

• If the data format doesn’t match expectations, it adjusts the cleaning logic.
• If the written output lands in the wrong position, it inspects the affected range and corrects the coordinates.

This small-step, feedback-driven approach greatly improves both ACCURACY and POSITIONING performance.

To keep long interactions efficient, UniverAgent uses an “adaptive message cache” that inserts cache checkpoints automatically, avoiding repeated computation over large, mostly stable context like sheet structure.

5. Online Architecture & Hybrid Execution

Unlike offline setups that automate desktop Excel or operate on static files, UniverAgent runs directly on the Univer online spreadsheet engine with a hybrid model:

• Python (pandas / numpy): Handles the “heavy” logic — multi-table joins, pivoting, advanced statistics. Python’s rich data ecosystem is fully leveraged here.
• JavaScript (Univer SDK): Handles the “fine-grained” operations — conditional formatting, column width adjustments, chart insertion, precise cell writes. All of this runs in a DOM-independent, headless environment, ensuring rendering consistency.

UniverAgent transforms complex spreadsheet tasks into deep data analysis and native-grade spreadsheet manipulation. As models get better at code, orchestrating Python plus JavaScript becomes increasingly powerful for complex tasks.

Because it operates on a live spreadsheet:

• State is preserved in memory, not shuffled through files.
• What-you-see-is-what-you-get: users can watch and supervise in real time.
• Version history makes agent mistakes easy to roll back.

6. Conclusion: Toward Autonomous Spreadsheet Engineering

A score of 68.86% on SpreadsheetBench is not just a number — it signals a transition in spreadsheet automation, moving from toy examples to production-grade capabilities.

UniverAgent demonstrates that with:

• strong context engineering
• a robust CodeAct loop
• and a native online spreadsheet engine

AI systems can reliably handle complex spreadsheet workflows.

This is no longer about generating auxiliary scripts — it’s about building autonomous spreadsheet engineers that understand intent, execute precisely, and correct themselves.

As model capabilities advance and tooling ecosystems mature, we expect to see meaningful adoption in enterprise environments, bringing spreadsheet automation closer to its full potential.

If you’re interested in the direction we’re building toward, feel free to reach out: developer@univer.ai