← back
Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs
Takeaway
Spreadsheet agents improve when persistent code execution and calculation or visual feedback replace fragmented tool calls.
Summary
- Witan Labs improved an internal financial spreadsheet benchmark from roughly 50% to 92% after exploring multiple agent architectures and representations.
- A rigid three-agent pipeline and standalone SQL or XML representations underperformed; partial tabular views and rendered images remained useful.
- Replacing about 15 tools with a persistent Node.js REPL enabled composable spreadsheet operations over a C# implementation.
- The REPL initially lifted accuracy to 74%; better search, formula tracing, prompts, and bug fixes drove later gains and effectively eliminated five-minute timeouts.
- Formula calculation and rendering engines provide feedback analogous to compilers and tests in coding workflows.
spreadsheetsreplfinancial-analysis
Original description
https://github.com/witanlabs/research...