Teaching Coding Agents to do Spreadsheets - Nuno Campos, Witan Labs

2026-07-08 · Source: AI Engineer · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

Nuno Campos from Witan Labs details their four-month effort to improve coding agents' proficiency in spreadsheets, boosting accuracy on a financial analysis benchmark from 50% to 92%. Initial attempts with multi-agent architectures and various spreadsheet representations (SQL, XML) proved rigid or ineffective. The breakthrough involved replacing numerous individual tools with a single Node.js REPL, leveraging JavaScript for agent interaction and C# for core spreadsheet logic. This REPL provided persistent state, enabling agents to build on prior work and interleave reasoning, leading to shorter, more effective scripts and eliminating timeouts. Key to success were high-fidelity formula and rendering engines, which provided crucial feedback loops for verification, alongside incorporating domain knowledge into prompts and employing deterministic evaluation strategies.

Key takeaway

For AI Engineers developing agents for complex, multi-step tasks like spreadsheet automation, you should prioritize implementing a REPL-based architecture with persistent state. This approach significantly boosts efficiency and accuracy by enabling agents to interleave reasoning and combine actions, reducing timeouts. Invest in building high-fidelity domain-specific feedback loops, such as calculation or rendering engines, to provide robust verification. Continuously refine your agent's interface and prompt-based domain knowledge for optimal performance.

Key insights

Persistent state REPLs and high-fidelity feedback loops significantly enhance coding agents' spreadsheet mastery.

Principles

• Replace sequential tool calls with a real scripting language.
• High-fidelity feedback loops are critical for agent verification.
• Domain knowledge in prompts guides agent focus.

Method

Implement a Node.js REPL with persistent state for agent interaction, backed by high-fidelity domain-specific engines for verification and feedback, and inject targeted domain knowledge into prompts.

In practice

• Adopt REPLs for complex multi-step agent tasks.
• Develop high-fidelity domain engines for agent verification.
• Utilize deterministic evaluation over LLM-as-a-judge when possible.

Topics

• Coding Agents
• Spreadsheet Automation
• REPL Architecture
• LLM Feedback Loops
• Financial Modeling
• Deterministic Evaluation

Best for: Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• Spreadsheet-RL: Advancing Large Language Model Agents on Realistic Spreadsheet Tasks via Reinforcement Learning — Spreadsheet-RL uses RL fine-tuning to create specialized LLM agents for complex, multi-step spreadsheet automation in Excel.
• Autonomous Long-Running Coding Agents — Autonomous coding agents require robust control systems with goals, evaluators, verifiers, and loops for long-running, verifiable task completion.
• How I Continually Improve My Claude Code — Continual learning for coding agents significantly boosts efficiency beyond out-of-the-box performance.
• Self-Improving AI Coding Agents Through Accumulated Behavioral Rules: A Closed-Loop Framework — Codifying human review feedback into persistent behavioral rules enables AI coding agents to self-improve across sessions.
• I built an open-source multi-agent SDLC harness that beats a cold Claude Code run on large repos, by learning the repo once. Real benchmarks (incl. where it loses) inside. [P] — Persistent knowledge bases for AI coding agents drastically reduce costs and improve efficiency on large codebases.
• I guess we're writing loops now? — Agents can dynamically orchestrate complex, multi-stage coding tasks through self-prompting loops, automating entire development workflows.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Engineer.