Excel Agent

Why most spreadsheet tooling is useless on real models

Libraries that read spreadsheets read values. A financial model is not values — it is a dependency graph of formulas, and the values are just its most recent evaluation. Read the values and you have a photograph of a machine.

Excel Agent works on the graph, and it does so from inside Excel: an Office.js taskpane add-in (plugin/manifest.xml, ReadWriteDocument permission) whose React workspace reads workbook context — cell snapshots, named ranges, sheet protection, charts and tables — through the Excel JavaScript API, and talks to a FastAPI backend over a typed API generated from the backend’s OpenAPI schema.

On the backend, formula_parser.py tokenizes formulas into an AST, builds the dependency graph, finds circular references by DFS cycle detection, and checks function arity against a table of Excel function specs. formula_graph.py turns that into a durable graph artifact — precedents, dependents, volatile functions, external links, broken refs — over analysis from a Node workbook-worker running SheetJS, ExcelJS and HyperFormula, so the graph can be recalculated, not just read. That is what answers the questions that actually come up in a model review: what feeds this cell, what breaks if this assumption moves, where is the hardcode buried in the middle of a formula column that should be uniform.

Auditing before authoring

The first useful behaviour is not writing formulas but finding the defects that model reviews look for by hand — inconsistent formulas across a row that should be uniform, constants embedded where a reference belongs, circularity, ranges that stop one row short of the data, sign convention breaks between sections.

Those are mechanical to detect on a graph and tedious to find by eye, which is exactly the right split of labour. Four verifier agents — syntax, logic, performance and QA — review every plan, and the parser runs financial sanity checks alongside the structural ones (WACC outside 5–15%, tax rate outside 0–50%, magnitude errors with auto-fix suggestions). The audit surface is held honest by a benchmark suite of thirty fixture workbooks — DCF terminal-value chains, LBO cash sweeps, merger accretion/dilution, three-statement models, broken-ref and spill-edge cases — run as a deterministic quality gate, with 232 tests across the backend, the add-in and the worker.

Editing without collapsing

When it does write, it writes formulas that stay formulas. An agent that helpfully replaces a model’s logic with the numbers that logic produced has destroyed the model while appearing to have updated it — and it looks correct until the first time someone changes an input.

The finance templates — DCF, LBO, three-statement, comps, merger, sensitivity, amortization — generate cell operations whose payloads are formula strings, not values: a projection row is emitted as =B19*(1+$B$5), so the model the agent builds recalculates like one a human built. The write path is just as conservative: the add-in applies operations atomically through the Excel JavaScript API, the apply gate blocks any write whose simulation artifact has not passed, and each mutation first captures a rollback snapshot of the range’s values, formulas and number formats. LLM routing is local-first by task complexity — LM Studio, then Ollama, escalating to Anthropic or OpenAI only when the task earns it — so routine edits never leave the machine.