Flow, well-being, and knowing when to stop
When an agent does the implementation, what is the developer absorbed in? How can we tell absorption from compulsive checking, and how might tools help people disengage?
About
Software developers are among the first knowledge workers to hand a large share of their work to AI agents. They now spend their days specifying, steering, and reviewing work that agents do in parallel. Industry tells this story in throughput: more output per person, faster delivery, more tasks in flight.
The people living through the change tell a more mixed one. They describe days spent reviewing more than making, attention stretched across several agents at once, skills they fear are fading, systems they answer for but no longer fully understand, and work that satisfies them less.
This workshop invites HCI, CSCW, and software engineering researchers, designers, and practitioners to take those experiences seriously. We ask what happens to developers' attention, creativity, and well-being when agents do much of the doing, what organizations should measure beyond throughput, and what we should keep for ourselves even when delegating would be faster. We treat software as an early case of a shift that other knowledge work will follow.
Themes
Five themes are starting points. Submissions may address any of them, and we expect participants to add others.
Attending to many agents at once
What sets the limit on how many threads a person can follow? How could tools batch, pace, or summarize what agents return? A delegated task keeps moving while one is away. What does that change about resuming work?
Understanding and owning work one did not do
What does it mean to understand, own, or answer for a system that agents wrote? How do people decide how closely to check a change? What happens to learning on the job, and to newcomers?
Creativity and what we keep for ourselves
Does working with agents make people more or less creative, and in which parts of their work? Which tasks should people keep because doing them is creative, exercises judgment, teaches them something, or gives the work meaning?
Measures beyond throughput
Once agents do the shipping, throughput says little about the person who directed them. What should complement it? We are interested in measures of attention and recovery, understanding, review effort relative to risk, learning, exploration of alternatives, and how people feel about their work.
Running through all five themes is a question about reach: what here is specific to code, and what will hold for writing, analysis, design, or research? We welcome participants from other knowledge-work fields who see the same pattern in their own work.
Call for participation
We invite three kinds of submission:
- Position papers of two to four pages in the ACM single-column format.
- Pictorials or other visual contributions.
- Practitioner stories of one to two pages describing your own or your team's experience of agentic work.
Contributions may report early findings, describe systems or design ideas, propose new measures, or argue a position. We welcome researchers, designers, and practitioners from HCI, CSCW, software engineering, and other knowledge-work fields. The organizers review submissions without anonymization for relevance, potential to spark discussion, and diversity of discipline, sector, seniority, and geography.
Accepted submissions will be posted on arXiv and on this website. They are not formal proceedings, so authors remain free to publish extended versions elsewhere. Authors may also bring a poster.
At least one author of each accepted submission must attend in person. Every participant must register for the workshop and for at least one day of CHI 2027. The workshop is open to all CHI attendees, with priority for accepted authors.
Submission instructions and the submission link will be posted here when the call opens.
Important dates
| Call for participation opens | December 17, 2026 |
| Submission deadline | To be announced, around February 2027 |
| Notification | To be announced, before CHI's early registration deadline |
| Workshop | CHI 2027, Pittsburgh, May 10–14, 2027 (day to be announced) |
Format
The workshop runs as two consecutive 90-minute sessions in person. Instead of a sequence of paper talks, it is built on Liberating Structures: short facilitation formats that give everyone in the room a voice and move quickly between individual reflection, small groups, and the whole room.
Session 1: What is happening to us?
The session opens with a one-minute pitch of every accepted contribution to the whole room. Participants then meet in quick paired conversations about a moment they brought from their own work or research, when working with agents made someone faster but worse off, and about the pitches they want to follow up. After two or three short invited talks, small groups name the questions that none of the pitches or talks answered.
Session 2: What should we do about it?
Groups first design the worst possible future for agentic work, then name the current practices that resemble it. The room then proposes and ranks research questions, design ideas, and measures, and drafts a shared research agenda.
After the workshop, the organizers and volunteers will write up the agenda and post it here, with a short version planned for ACM Interactions.
Every activity that involves moving around the room has a seated version. All prompts are shown on the projector and on printed cards, and outputs are captured in a shared document. Papers, poster PDFs, and slides stay on this website for anyone who cannot attend.
Organizers
- Shyam AgarwalCarnegie Mellon University, USA
- Kelly BlincoeUniversity of Auckland, New Zealand
- Andrew MacveanGoogle, USA
- Daniel RussoAalborg University, Denmark
- Bogdan VasilescuCarnegie Mellon University, USA
Questions: vasilescu@cmu.edu