Nuno Campos is co-founder and chief technology officer of Witan Labs, creator of LangGraph, and a founding LangChain engineer. His work centers on a demanding question for autonomous software: how to make AI agents reliable when their tasks require persistent state, specialized knowledge, and results that can actually be checked.
Before joining LangChain as its first employee, Campos contributed to open-source JavaScript tooling, including the Enzyme testing ecosystem. At LangChain, he led open-source engineering and created LangGraph, a framework for stateful agents designed around durable execution, human intervention, and explicit developer control. With Mayo Oshin, he co-authored Learning LangChain, published by O’Reilly in 2025.
At Witan Labs, Campos applies those concerns to financial spreadsheets, whose meaning depends on visual layout, formulas, assumptions, formatting, and dependencies that disappear when workbooks are flattened into ordinary text. Witan’s open-source spreadsheet toolkit gives coding agents capabilities for editing workbooks, calculating formulas, rendering sheets, and checking their work.
The ideas behind his agent architecture
- Give agents a persistent programming environment. After experimenting with rigid multi-agent workflows and standalone SQL, XML, CSV, and HTML representations, Campos consolidated approximately fifteen spreadsheet tools into a persistent Node.js REPL. Agents can compose operations in JavaScript, preserve variables between calls, and alternate short programs with fresh reasoning; a C# engine handles spreadsheet execution, while TypeScript definitions describe available capabilities. Witan’s internal financial-analysis benchmark improved from roughly 50% to 74% after introducing the REPL and eventually reached 92% with additional refinements, including formula tracing and search. These figures reflect an internal evaluation, not an independent industry benchmark.
- Build verification systems that match the real domain. Campos treats spreadsheet calculation and rendering as counterparts to a compiler and test suite: agents need trustworthy ways to inspect formulas, layout, and outputs before correcting mistakes. Incomplete engines can turn valid spreadsheet operations into apparent failures, making high-fidelity verification loops more important than any particular agent interface. He expects interfaces to change as model capabilities evolve; the need for accurate feedback remains.
- Define correctness with an authoritative oracle. In his essay on loop-driven development, Campos argues that inexpensive AI-generated code and tests make the definition of correct behavior the harder problem. His approach to loop-driven development uses external, deterministic authorities—such as Microsoft Excel itself—to generate expected outcomes and reproducible tests. He favors comparing generated spreadsheets with reference workbooks where possible, using model-based judging when deterministic evaluation cannot capture the task.
- Design the whole agent, not merely its tool connections. Campos has challenged claims that interoperability protocols alone can make agents production-ready, arguing that prompts, architecture, domain expertise, and execution interfaces must fit the work together. Witan’s public engineering log documents the practical stakes: an individual-command spreadsheet workflow underperformed openpyxl on an internal comparison because repeated process launches and a recalculation bug undermined otherwise capable models.
