# Let agents edit Office, and check the results

> Connect Univer’s Office tools to the agents you already run: read, edit and calculate Office content through an API or CLI, then check and fix results from values and renders.

Editing, calculation and checking tools for the agents you already run. No desktop Office, and everything stays editable.

<https://univer.ai/agents/>

Read → Edit & calculate → Read back → Verify & fix

## Plug Office tools into your agents

Your harness keeps its models, orchestration and Skills; Univer provides the Office tools. Agents call them through an API or CLI, get results and feedback back, and carry on.

## Understand, create and edit Office content

Agents locate and change content and run calculations, producing native Office content people keep editing.

## Check the result, fix from feedback

Tools return edit status, calculated values and renders. Agents check them against the task, fix what’s off and verify again.

## Checks you can see: render and layout

Beyond values, agents get page renders and layout checks. When a slide title runs off the page, the check flags it; the agent fixes it and renders again. You decide what passes.

## Ranked #1 on SpreadsheetBench

SpreadsheetBench accuracy, pass@1, December 2025 ([Evaluation notes](https://univer.ai/#:~:text=SpreadsheetBench))

|  | % |
|---|---|
| Human | 71.3 |
| **Univer** | 68.86 |
| Bluebox | 62.9 |
| Shortcut.ai | 59.25 |
| Copilot in Excel (Agent Mode) | 57.2 |