Where an analyst sees a revenue table and a P&L summary, an LLM sees 10,000 cell values and nested SUMPRODUCT formulas. Over four months we pushed our agent's pass rate from 50% to 92%, through several architectures, many dead ends, and one decision that mattered more than the rest: replacing fifteen discrete tools with a single persistent Node.js REPL.
This talk is a field report on what worked and what didn't. The REPL lets the agent write JavaScript instead of stringing tool calls together, with variables that persist so it can explore, reason, and course-correct mid-task. Underneath sits a verification loop — formula calculation, rendering, linting - that the agent uses to check its own work, and that has compounded across three frontier model releases rather than been made redundant by them.
The lessons generalise: interfaces are ephemeral but verification engines are durable, domain knowledge outlasts tooling (we changed backends four times; the financial-domain prompt carried over every time), evaluation is roughly half the work, and most "agent confusion" is really a plumbing problem
Nuno Campos is Co-Founder and CTO of Witan Labs, building headless tools for autonomous agents. Previously he was the first employee at LangChain, creator of LangGraph, author of Learning LangChain.
