AI Engineer host Yijia Liang, Compound AI co-founder and former Coatue vice president, said his team rebuilt Excel, PowerPoint and Word from scratch to run long finance jobs in parallel without touching desktop Office.
That is not a plugin.
Liang said the rebuild was a response to how buy-side deal teams actually work. Liang argued personal agents took off because they handle reliable background tasks, act proactively, and plug into email and calendar, but that breaks in institutional finance where a single data room can hold 15,000 files and a firm may sit on petabytes in SharePoint or Box. He said most firms today run single-player ChatGPT or Claude workflows, then download files and circulate them on email or Teams for review. Compound made every agent come with a built-in inbox so an entire deal team can talk to the same agent on the same thread, Liang said, and the system has to authenticate read and write access on each thread because every deal is under NDA. He added the team deliberately adds latency when multiple people email quickly so the agent does not spam acknowledgements and feels more human.
For execution, Liang said desktop Excel, PowerPoint and Word are bound to a single CPU and existing engines were not enough. The company tested desktop plugins, virtual machines for computer use, and both open-source libraries like OpenPyXL and commercial engines, then decided to recreate Office with agents in loops using Office as an oracle. Liang said the team started from the open-source Iron Calc for Excel and, after 1.5 trillion input-output tokens and more than 8 million lines of code, with about 800,000 lines of Rust for the Excel engine alone, produced what he described as performant, one-for-one replicas for Excel, PowerPoint and Word rendered in Rust for pixel-perfect parity. Liang said a headless engine recalculates 20 MB Excels about 4x faster and lets models, slides and memos build in parallel, and because Compound owns the write, render, check and fix loop it can run iterative correction without guessing at formatting. The identical rendering matters because firms only trust what they can download and send, he said.
On context, Liang split the problem into breadth and depth. For breadth, he said Compound prefers agent swarms over indexing everything into a vector database, with a lead agent slicing a file directory and swarms reading files via grep or similar methods and returning cited findings grounded in source documents. For depth, he said a 30 MB workbook with 2 million formulas and five sensitivity tables that takes five minutes to recalc chokes off-the-shelf models that print the whole context into code via OpenPyXL, while Compound's own engine exposes richer APIs that make processing faster and cheaper for the model. That stack, from connected sources like SharePoint and Outlook to cited findings to firm-specific preferences for how a model should look, is meant to leave humans imparting judgment on whether a deal is good rather than checking if numbers tie, Liang said. He framed the shift as moving finance from copilot to autopilot, where a 50-page investment committee deck or a full IC model can be delegated from a data room drop in days instead of four weeks of iterations. The pattern matches what the AI PM Wiki describes as Compound Engineering, a workflow for managing goal-driven agent work through structured build loops.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 11,995 subscribers.
Liang demoed the workflow by building his own talk deck in Compound with comments driving formatting, and said a team of five can rip through a 100-page deck in an afternoon. He declined to detail pricing, deployment requirements, how the inbox authentication is audited for NDA enforcement, or what happens when the headless render diverges from Microsoft Office on an edge case. Liang did not say who owns liability when an autopilot memo or model misses a cited number after the firm has scaled from seeing a handful of deals to ten times that volume, and that is the question deal teams will have to answer before they let an agent run the book.
