Reproduce “Auditing and Controlling AI Agent Actions in Spreadsheets” on OpenResearch

About this paper

Sadra Sabouri, Zeinabsadat Saghi, Run Huang, Sujay Maladi, Esmeralda Eufracio, Sumit Gulwani, Souti Chattopadhyay

Researchers from the University of Southern California and Microsoft introduce Pista, an AI agent for spreadsheets designed to enhance human oversight by integrating verification and control directly into the execution workflow. A usability study found Pista improved users' comprehension of agent actions and their ability to steer agent behavior through step-by-step transparency and granular intervention, leading to a stronger sense of control and co-authorship.

Key insights

• Active human participation and oversight *during* AI agent execution, rather than solely post-hoc review, is critical for achieving meaningful human-AI collaboration and fostering calibrated trust.
• Integrating verification and steering capabilities directly into the agent's workflow significantly enhances user comprehension, improves their sense of control, and promotes co-authorship over the final output.
• The concept of a 'semantic diff' — surfacing underlying computational logic, dependencies, and scope of change — is a fundamental design primitive for making agent reasoning legible and verifiable across artifact-producing agents.

Summary by alphaXiv.