Yijia Liang, founder of Compound and a former finance analyst, describes the company’s approach at AI Engineer.

A shared teammate

  • Deal teams need shared context and review across people, not isolated chat sessions.
  • Compound gives each agent an email inbox; Liang says access checks and batching replies are important for confidential threads.
  • A data room can contain thousands of files, while large workbooks contain complex dependencies.

An engine for the work

  • Compound evaluated Office plugins, desktop automation and existing libraries before building its own document engines.
  • Liang reports spending 1.5 trillion input and output tokens and producing over eight million lines of code across the effort.
  • The spreadsheet engine builds on the open-source IronCalc project; Office served as a comparison oracle for iterative tests.
  • The team claims faster recalculation and faithful rendering. These are company claims, not independently verified compatibility or performance results.
  • Owning the engine supports a write, render, check and fix loop and lets teammates leave comments for the agent.

Finding context

  • Liang describes a lead agent assigning groups of files to other agents and returning cited findings.
  • A specialized workbook engine can expose a more structured view of a large spreadsheet to the model.
  • Citations help people inspect the source, but they do not establish that a financial analysis is correct.

“How can they all talk to the same agent so that one agent can share the context of everyone who matters for any particular work stream?”