ACE: Moving toward Co-Investigation with the Agentic Code Explorer

Generative AI models are commonly used today to aid in co-creative tasks in which a human user and a generative AI model are able to collaboratively modify an artifact-under-creation, such as a piece of writing, source code, an image, or a musical composition. The focus of this interaction is on the process of co-creation, where the end goal is to produce an artifact that satisfies the user's needs and (under some circumstances) the organization's needs. In this paper, we argue that the emergence of the agentic design pattern, in which an LLM-based agent is capable of invoking external tools and iteratively refining its own outputs, pushes the potential value of generative models beyond mere co-creation (where generated artifacts are the end goal) into a space of co-investigation (where generated artifacts are the means that facilitate a broader conceptualization, learning, or problem solving process).

As a step toward this goal, we present ACE -- the Agentic Code Explorer -- a prototype agentic system for helping software engineers conduct sensemaking tasks within large code repositories. ACE serves as a testbed for examining new approaches to human-AI collaboration for co-investigative tasks with a specific focus on incorporating features that help users establish a mutual theory of mind with the agent.