A talk by Nuno Campos from Witan Labs covering 4 months of work improving coding agents for spreadsheet tasks, going from 50% to 92% accuracy on a financial analysis benchmark. Key findings include: replacing 15 individual tools with a Node.js REPL as the primary agent interface dramatically improved performance (50%→74% alone), persistent REPL state enabled shorter iterative scripts with more reasoning interleaved between steps, building high-fidelity formula and rendering engines created a reliable verification feedback loop, and deterministic evaluation outperforms LLM-as-judge where possible. Dead ends included multi-agent architectures with rigid discovery phases and various spreadsheet representations (SQL, XML, CSV, HTML). Generalizable lessons: give agents a real scripting language instead of many sequential tool calls, invest in domain-specific feedback loops, add domain knowledge to prompts, and always inspect traces to distinguish model errors from infrastructure bugs.