A Team of One: How Individuals Orchestrate Multiple AI Agents at Work
Motivation & Problem
A new way of working is emerging at the edges of practice. Individual knowledge workers — freelancers, developers, consultants, analysts — no longer use a single AI assistant to help with a single task. Instead, they set up several AI agents at once: delegating work, letting agents run in parallel with limited supervision, reviewing what comes back, and stitching the results together. A frequently voiced expectation is that this becomes the normal shape of knowledge work: that many workers will, in effect, direct and coordinate a set of agents rather than execute the work themselves.
If that is true, it is a remarkable shift. Delegating and coordinating work has historically been managerial work, and how much of it one person could do was bounded — not by ambition, but by the cost of coordinating and checking the work of others. Agentic AI appears to lift some of these limits while imposing new ones. Anecdotally, practitioners can now generate far more work than they can meaningfully review, so the bottleneck shifts from producing output to verifying it, and from doing the work to keeping track of it.
We know very little about how this actually happens. How do people decide what to hand to an agent and what to keep? How do they check work they did not do? How do they keep context, quality, and responsibility intact across several parallel streams? And what, in practice, limits how many agents one person can effectively direct? This thesis is an early, exploratory study of these questions with people who are already working this way. The overarching question is: How do individuals orchestrate multiple AI agents in their work, and what bounds the scope and number of agents they can effectively direct?
The study is deliberately exploratory, but not merely descriptive: rather than collecting a list of practices and pain points, the aim is to arrive at a framework, typology, or process model that explains how orchestration is accomplished and what constrains it.
What You Will Do
- Review the literature on human–AI collaboration and agentic AI in information systems, management, and HCI/CSCW, and identify suitable sensitizing concepts (e.g., span of control).
- Design the study. Develop an interview guide and a sampling strategy targeting people who already orchestrate multiple agents — freelancers, developers, agencies, and employees in early-adopting organizations.
- Conduct interviews. Run semi-structured interviews, ideally combined with walkthroughs of participants’ actual setups and critical incidents, so that accounts stay grounded in concrete work rather than opinion.
- Analyze the data using an established qualitative approach.
- Develop and write up a framework of orchestration practices and their constraints, discussing implications for how knowledge work, expertise, and responsibility are organized when one person directs many agents.
Your Profile
- Genuine interest in qualitative, interpretive research and in how AI is changing knowledge work.
- Comfort with reaching out to practitioners, conducting interviews, and listening carefully — the quality of this thesis depends on access and on good conversations.
- Patience for systematic, iterative coding and analysis; a structured and independent working style.
- Nice to have (not required): own experience with AI agents (e.g., coding agents or agentic workflows), or existing contacts in relevant communities.
Please get in touch with a short email including your CV, a current transcript of records, a few sentences on why this topic interests you, and the planned start/finish date: hise@ifi.uzh.ch