Nuno Campos describes unsuccessful attempts to represent spreadsheet work through a rigid multi-agent process or a single text format. CSV views and rendered images remain useful components, but his team found that agents needed ways to explore structure, formulas and dependencies as their understanding developed.
Nuno Campos explains how a persistent JavaScript REPL lets agents combine operations and retain variables across calls. He reports an improvement from about 50% to 92% on an internal financial-analysis benchmark after this change and later refinements. Those figures describe the team's benchmark, not independently verified accuracy across all spreadsheets.
Nuno Campos emphasizes calculation and rendering as feedback loops whose value depends on faithful underlying engines. He also describes replacing model-based grading with deterministic input-output comparisons where possible and inspecting traces for tool, prompt or infrastructure bugs that can resemble reasoning failures.
Watch on YouTube
