How Much AI Agent Code Merges Without Human Review

Chart of agent-authored code reaching main without human diff review at 14 engineering teams: 11 reported 0%, one 0 to 5%, one 20% and one about 100%

AI-generated code review stayed with people on 11 of 14 engineering teams, which meant 0% of agent code reached main unreviewed. The median across all 14 teams was 0%.

Study methodology: In August 2026, 40 practitioners responded to a public journalist query, each describing 1 live AI agent deployment. 14 worked in engineering and reported how much agent-authored code reached main without human diff review. Most respondents were founders or heads of function at firms of up to 31 people. The figures, therefore, describe the stated policy on small teams.

Essential statistics: AI-generated code review before merge

  • 11 of 14 engineering respondents (79%) reported that 0% of agent code reached main without a person reading the diff.
  • The median share of unreviewed agent code across all 14 respondents was 0%.
  • 2 respondents stayed at or below 20%: an enterprise architect reported 0-5%, and Pikadil reported 20%.
  • Wolf Capital Japan was the single outlier, merging about 100% of agent code without diff review.
  • Engineering had the lowest mean autonomy among the 4 functions, at 2.31 out of 5 across 13 deployments.
  • Faros AI’s 2026 telemetry from 22,000 developers found that, with high AI adoption, 31.3% more pull requests were merged unreviewed.

Key takeaways

  • AI-generated code review is held at the merge step on most teams, with 11 of 14 reporting a flat 0%.
  • The report reads the 0% as a gate held deliberately in front of high agent adoption, not as a sign of light agent use.
  • Stated policy and fleet telemetry diverge. On teams this small one person can read everything, while Faros recorded more unreviewed merges under high AI adoption, where volume outran reading capacity.
  • On teams this small, the reviewer is often the founder or the only engineer, so review depends on 1 person’s time.
  • The 3 exceptions range from 0 to 5% up to about 100%, and 1 team accounts for the entire high end.

Actionable insights

  • Require a human diff review before agent code reaches main, as 11 of 14 engineering respondents do.
  • Measure AI-generated code review from repository data alongside the written policy, since Faros found 31.3% more unreviewed merges under high AI adoption.
  • Plan reviewer hours before raising agent output, because review on small teams often rests with 1 engineer.
  • Grant authoring and merge rights to agents separately, so an agent can write a change while a person holds the merge.

“Agents write most of the code, and the share that reaches main without a human reading the diff is zero. That human is me.” — Ivan Cordeiro Junior, Founder, Marfin

Agentic AI in Production: SumatoSoft study of 40 live AI agent deployments
Agentic AI in Production: Where Agents Run Unattended and Where They Get Pulled Back
View research
SumatoSoft logo
If you have any questions, email us info@sumatosoft.com

    Please be informed that when you click the Send button Sumatosoft will process your personal data in accordance with our Privacy notice for the purpose of providing you with appropriate information.

    Vlad Fedortsov, IT Sales Expert at SumatoSoft
    Vlad Fedortsov
    Account Manager
    Book an intro call
    Thank you!
    We've received your message and will get back to you within 24 hours.
    Do you want to book a call? Book now
    SumatoSoft clients logo