Abstract

As AI agents become increasingly embedded in software development teams, developers are no longer simply using AI but are beginning to collaborate with these AI agents as teammates. This shift raises a critical question: do developers trust these agents enough to work alongside them, and under what conditions are they willing to delegate meaningful tasks such as code generation, debugging, and system design? While prior research on AI in the workplace largely treats AI as a tool rather than a teammate, yet this distinction matters profoundly. Tools are operated; teammates are trusted (Baird & Maruping, 2021). Unlike traditional software development tools (e.g., IDEs or automation scripts), modern AI agents do more than merely execute predefined commands. They interpret requirements, generate and refine code, suggest architectural decisions, and adapt outputs based on evolving project contexts. In doing so, they assume responsibility for tasks with ambiguous requirements and seek optimal outcomes under uncertainty (Dennis et al., 2023). As a result, understanding how trust forms in human–AI collaboration—and how it translates into delegation decisions—has become essential. Software development tasks are interdependent, iterative, and often involve varying levels of requirement ambiguity, making delegation decisions inherently uncertain. Consequently, outcomes in Human-Agent Teams (HATs) are shaped by both the task characteristics and developers’ perceptions of AI agents’ intelligence and trustworthiness (Dolatkhahi & Li, 2025; Iftikhar et al., 2024). However, prior research has yet to fully explain when and how developers’ perceptions of AI agent intelligence translate into trust, and how such trust subsequently drives delegation decisions across different development tasks. Grounded in HAT literature and trust literature (e.g., Iftikhar et al., 2024; Hoesterey & Onnasch, 2023), this study examines the mechanisms through which developers’ perceptions of AI agent intelligence shape their trust in AI teammates and, in turn, influence delegation behavior. We conceptualize perceived AI agent intelligence as a multidimensional construct encompassing perception, comprehension, action, and learning capabilities (Bawack, 2021), enabling AI agents to function as adaptive collaborators in iterative development processes. We also propose a moderate mediation model in which these relationships are moderated by three factors, specifically, perceived task risk (high vs. low), task requirement clarity (high vs. low), and collaboration structure (human-led, agent-led, co-led). To test our model, we will conduct a randomized field experiment in software development teams across organizations actively deploying AI development agents. This study will contribute to theory by reconceptualizing AI from tool to teammate in software development context and by identifying the conditions under which trust forms and translate into delegation behavior across different task and collaboration contexts. For practice, this study is expected to provide guidance on how organizations can structure AI delegation in software teams, particularly by aligning task characteristics with appropriate levels of trust. It also offers insights into how collaboration structures may be designed and sequenced to support effective human–AI teamwork.

Download

Share

COinS