TableTalk: Scaffolding Spreadsheet Development with a Language Agent

• DOI: 10.48550/arxiv.2502.09787
• Publication Date: 2025-02-13
• Authors: Jenny T. Liang; Aayush Kumar; Yasharth Bajpai; Sumit Gulwani; Vu Le; Chris Parnin; Arjun Radhakrishna; Ashish Tiwari; Emerson Murphy-Hill; Giulia Lisboa Car...
• Keywords: Software Engineering (cs.SE); FOS: Computer and information sciences; Computer Science - Software Engineering; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computer Science - Human-Computer Interaction; Human-Computer Interaction (cs.HC)

Despite its ubiquity in the workforce, spreadsheet programming remains challenging as programmers need both spreadsheet-specific knowledge (e.g., APIs to write formulas) and problem-solving skills create complex spreadsheets. Large language models (LLMs) can help automate aspects of this process, recent advances planning reasoning have enabled agents, which dynamically plan, use tools, take iterative actions complete tasks. These agents observe, act, making them well-suited scaffold by following expert processes. We present TableTalk, a agent that helps build spreadsheets conversationally. Its design reifies three principles -- scaffolding, flexibility, incrementality we derived from two studies seven 62 Excel templates. TableTalk structures development generating step-by-step plans suggesting next steps users choose from. It also integrates tools enable incremental construction. A user study with 20 shows produces 2.3 times more likely be preferred over baseline agent, while reducing cognitive load time spent about 12.6%. TableTalk's approach has implications for human-agent collaboration. This includes providing persistent direct manipulation interfaces stopping or undoing actions, ensuring such accepting deactivated.