Teaching Coding Agents to do Spreadsheets
summary
Started at ~50% accuracy on a financial benchmark and reaches 92% accuracy after swapping a 15‑tool pipeline for a single Node.js REPL that runs short JavaScript sessions with persistent state while heavy lifting stays in C#. Built a high‑fidelity formula engine and render engine so the agent can verify outputs; result: near‑zero 5‑minute timeouts, fewer long tool chains, and much faster, more reliable spreadsheet edits.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 1:57Spreadsheets and AI challenge
- 3:05Dead ends in representation
- 4:28Node.js REPL breakthrough
- 6:33REPL efficiency gains
- 8:22Results to 92% accuracy
- 9:48Verification loops needed
- 11:32REPL vs future models
- 12:37Domain knowledge in prompts
- 13:43Evaluation challenges
- 15:52Debugging agent failures
- 16:12Key takeaways
related talks
30mHow AI Agents Will Change the Future of Programming
45mAI agents in practice - beyond the prototype
48mAI Coding Agents and How to Code Them
16mCoding Reinvented How AI Agents Are Shaping the Future of Software Development
26mFrom autocomplete to agents: AI coding assistance state of play
30mBeyond the Hype: Using AI Coding Agents Today
07h 47mAI Agents Conference 2026 - Day 5
02h 04mShip Real Agents: Hands-On Evals for Agentic Applications