Europe 2026
Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs
Overview
This talk details the process of training AI coding agents to perform tasks within spreadsheets, aiming to match their proficiency in programming languages. The project began with a 50% accuracy rate on financial analysis benchmarks and ultimately achieved 92% accuracy. The presentation covers the challenges of representing spreadsheet data to AI, various approaches explored, and the breakthroughs that led to significant improvements.
Who should watch
- AI engineers working on agent development
- Product managers building AI-powered tools
- Developers looking to integrate AI into spreadsheet workflows
- Anyone interested in improving agent capabilities beyond traditional coding tasks
Key takeaways
- Spreadsheets are deceptively complex for AI due to their visual and implicit structure, requiring agents to infer meaning beyond raw data.
- Early attempts using multiple agents or rigid multi-step processes proved too inflexible, hindering context flow and iterative refinement.
- Various data representations like SQL and XML were explored but were not effective as standalone solutions, though CSV/TSV views and HTML formatting offered partial benefits.
- A major breakthrough was replacing numerous individual tools with a single Node.js REPL, allowing agents to combine multiple operations within one call and maintain persistent state.
- The REPL approach, leveraging JavaScript for its sandboxing and LLM familiarity, significantly reduced timeouts and improved efficiency compared to sequential tool calls.
- Implementing feedback loops, such as formula calculation and rendering engines, is crucial for enabling agents to verify and correct their work, similar to compilers and linters in coding.
- Deterministic evaluation methods, where possible, are more trustworthy than relying solely on LLM-based judgments for assessing agent performance.
- Adding domain-specific knowledge to prompts helps focus agents on relevant concepts like revenue or ARR, improving performance without extensive retraining.
Notable quotes
*Spreadsheets are a little bit harder for AI than you might think at first.*
*If your agent is making many sequential tool calls or even parallel tool calls, then you've kind of invented a bad scripting language.*
*Feedback loops really matter. And if you're happen to be working in the domain where those where you can build those feedback loops with existing tools, then great.*
Unofficial community note. Prefer the recording for nuance.