The authors adapted the Thought-Action-Result trajectory encoding to include human interventions, defining Human-AI loops that pair developer feedback with the agent's subsequent trajectory.
Over half of Copilot PRs are fully autonomous, and human-agent collaboration is typically brief
Analysis of 7,581 real-world Copilot-authored pull requests reveals that most work occurs in the initial trajectory, with developer interventions being focused and short.
Academic
Izah Sohail · Yerdanat Kurmantayev · May Mahmoud · Sarah Nadi
New York University Abu Dhabi
Research Digest··2 min read
The authors studied 7,581 Copilot-authored pull requests (PRs) from 1,602 GitHub repositories to examine human-in-the-loop agent trajectories.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.