RESEARCH · SOLO WORKFLOW
Rochester Institute of Technology researchers Maliha Noushin Raida and Daqing Hou analyzed 25,264 agentic pull requests across 2,361 popular GitHub repositories, Copilot, Codex, and Claude Code, created between May and July 2025. Their finding: human-agent collaboration is dominated by a single reviewer, and small teams that generate the most agentic PRs almost never build the multi-person review process that would let that output actually scale. This is a small working demo of both halves of that paper, the real collaboration-pattern breakdown and the backlog it produces once agent output outruns one person's review pace.
Real data · how 25,264 agentic PRs were actually reviewed
Every bar below is the paper's own measured share. Click a segment for what it means.
One human both reviews and commits the agent's contribution, alone. 19,488 of 25,264 agentic PRs, the single most common pattern in the dataset by a wide margin.
88.7% of every agentic PR in the dataset was handled by a single human, either reviewing and committing, or reviewing alone. Only 11.3% involved a second person at all.
Backlog calculator
The paper's reference productivity line is 36 agentic PRs per participant over three months, about 3 per week of real review capacity. Only 25 of 2,361 repositories (1%) crossed it. Set how many PRs your agent could generate per week and how many reviewers you actually have, and watch the backlog.
At 10/week generated against 3/week reviewed, unreviewed PRs pile up to 56 by week 8. The agent did not slow down. The reviewer did not speed up.
WHY THIS MATTERS FOR CLIENT WORK
I am exactly the profile this paper describes: one person reviewing and merging everything an agent proposes across my own and client repos, no second reviewer to split the load. The paper's numbers say that setup is not a phase I will grow out of as agents get better, it is the dominant pattern even in the small teams that generate the most agentic PRs. The agent was never my constraint. My own review pace is. The useful move is not asking an agent to write more, it is protecting review time as the actual scarce resource and being honest about how many client repos one person can responsibly hold open at once.
Built 24 July 2026 · data sourced from Raida & Hou, RIT (2026)