Agents in the Grid — Harsh Virani

Abstract

Spreadsheets are where office work actually happens, yet most agents treat them as text. This thesis takes the agent harness from coding agents to spreadsheets: a defined set of spreadsheet actions, run against a real spreadsheet program in an isolated container. The plan is to compare about five open-weight models on one harness, then fix the best model and compare my harness against ready-made agent harnesses, on SpreadsheetBench. I am in the literature review; nothing is implemented yet and there are no results.

01Why spreadsheets

Ask an agent to “reconcile last quarter's invoices” and it has to find the right sheet, read ranges it can't see all at once, respect formulas someone else wrote and leave the file in a state a human trusts. That's research and execution in the same loop.

A benchmark that only checks the final answer misses how the agent got there. In a spreadsheet the path matters too: a correct number written over a formula is still a broken file.

check · D3 · 5/5

Fig. 1 — the agent's loop on a sheet: look, read a range, reason, act, check

02The planned harness

The idea is to bring the agent harness from coding agents to spreadsheets. The agent gets a defined set of spreadsheet actions, such as reading cells, writing a formula, filtering, sorting and building a pivot table, and works on a real spreadsheet inside an isolated container. The current plan is LibreOffice with its Python bridge, one Docker container per task, no network.

# a sketch of the planned action set (not implemented)read_cells(range)write_formula(cell, formula)filter(range, condition)sort(range, key)build_pivot(source, rows, values)validate()

Planned actions, a sketch

A correct number written over a formula is still a broken file.

03The planned study

Two experiments, both on SpreadsheetBench ¹. First, one harness and about five open-weight models. Second, the best model held fixed while the whole harness is swapped: mine against ready-made agent harnesses, SWE-agent ² among the candidates. From the two together I want to argue whether a strong model or a strong harness matters more.

04Where it stands

I am in the literature review and shaping the study design with my supervisor. Nothing is implemented yet and there are no results. This page will update as the work moves.

References

1. Z. Ma et al. SpreadsheetBench: towards challenging real-world spreadsheet manipulation. 2024.
2. J. Yang et al. SWE-agent: agent-computer interfaces enable automated software engineering. 2024.