Skip to content
Primary sourcePodcastHarness atlasAutonomy boundariesNext review
Research Insights Made Simple
Software factories · series 1/8Series prologue

Over 40% of PRs are agent-started. Who accepts?

The dark factory: autonomy levels and what the numbers show

/ Research Insights Made Simple · Software factories 1/8 · Autonomy levels

Slide contents

  1. 1. Over 40% of PRs are agent-started. Who accepts?

    The dark factory: autonomy levels and what the numbers show

  2. 2. Agents write, people still accept

    Jellyfish, September 2026: median and p90 across companies

  3. 3. Check Shapiro’s ladder against data

    Frame, measurements, cases and the new bottleneck

  4. 4. Four tasks, four different rungs

    T1–T4 from the series; no runs, a teaching layout

  5. 5. Many autonomy scales, different yardsticks

    Shapiro measures the developer’s role

  6. 6. Six rungs from autocomplete to factory

    L0–L5, by analogy with driving automation

  7. 7. The rung changes the human’s job

    At L3 you read diffs; at L4, test reports

  8. 8. “90%” is an estimate, not a measurement

    No method stated; we test it against 2026 data

  9. 9. Three denominators, three different answers

    A platform, a vendor’s customers, company self-reports

  10. 10. Over 40% of PRs are agent-started

    GitHub: 17% in March; the signal is a floor

  11. 11. Median company: agents touch 27.7% of PRs

    September 2026; p90 at 57.9%, median ≈2% in January

  12. 12. Real factories are still rare

    Factory-opened PRs: p90 0.0%, p99 9.4%

  13. 13. Half a percent merges without humans

    AI-only review: median 0.5%, p90 8.5%

  14. 14. Stripe: agents write, humans accept

    Over 1,300 weekly PRs with no human-written code

  15. 15. The share grows, review remains

    Ramp: three in four PRs; Cursor: more than half

  16. 16. Spotify: auto-merge for a narrow class

    Maintenance after checks; review where it matters

  17. 17. StrongDM: no human code, no human review

    Rung five per Kahana; three engineers started it

  18. 18. The bottleneck moved to review

    Faros 2025: teams with high AI adoption

  19. 19. Change volume outran some checks

    Spotify’s review: pace, not AI-authored code

  20. 20. Kahana: shared blind spots undermine trust

    Horthy: tests miss the cost of the next change

  21. 21. Industry shares measure authorship, not quality

    Samples, floors, self-reports, correlations

  22. 22. Choose rungs per task, not per team

    Authorship and acceptance are separate axes