SheetMind's Three-Agent System Pushes Spreadsheet Automation Past 50% Accuracy

A team of researchers has built a spreadsheet assistant that turns plain-English instructions into working formulas and formatted data, and it does so with a level of reliability that outpaces its closest predecessor by a wide margin. Called SheetMind, the system completed 100% of the tasks it attempted on the SheetCopilot Benchmark while producing the functionally correct result 54.8% of the time. That correctness rate is a substantial jump from the 44.3% posted by SheetCopilot, the benchmark's namesake tool and the previous reference point for LLM-driven spreadsheet automation.

The benchmark itself is a 221-task suite designed to stress-test how well an AI system can interpret natural-language spreadsheet requests and execute them correctly, not just attempt them. SheetMind was tested using GPT-3.5-Turbo as its underlying language model, a relatively modest choice compared to more capable frontier models, which makes the accuracy gain more notable: the improvement appears to come from architecture rather than raw model horsepower.

How the three agents divide the work

SheetMind's design splits the automation task into three specialized roles. A Manager Agent takes a user's request, however complex, and breaks it down into a sequence of smaller subtasks. An Action Agent then converts each subtask into structured, executable commands, using a Backus-Naur Form grammar to constrain the output into a syntax the spreadsheet software can reliably parse. Finally, a Reflection Agent checks the generated actions against the user's original intent before anything is finalized, acting as a validation layer meant to catch mismatches between what was asked and what the system is about to do.

This division of labor is the likely explanation for the gap between execution success and functional correctness. Getting a command to run without error is one problem; getting it to do what the user actually meant is another, and it's the second metric where most spreadsheet copilots have struggled. By dedicating a distinct agent to intent-checking, SheetMind appears to close some of that gap, even if 54.8% still leaves plenty of room for improvement.

From research paper to Workspace extension

The team didn't stop at benchmarking. SheetMind has been integrated into Google Sheets as a Workspace extension, moving it from a purely academic exercise toward something usable inside a mainstream productivity tool. That step matters for adoption: spreadsheet automation research has a long history of impressive benchmark numbers that never make it into a product people can actually open and use.

The underlying paper first appeared on arXiv on June 14, 2025, with a revised version posted just over a year later, on July 13, 2026. As with most LLM-agent research, the numbers reflect performance on a fixed benchmark rather than open-ended real-world use, and results with more capable models than GPT-3.5-Turbo remain to be seen. Still, a multi-agent architecture that meaningfully lifts functional correctness on a standardized spreadsheet benchmark, and ships as an actual extension, is a rarer combination than the research literature usually produces.