Coding is solved. Its control is not.

The State of AI Code Control 2026 - The CiSCiN Report

112M

commits analyzed

35,000+

organizations, 159 countries

69%

use AI agents in their code work

75%

of agent pull requests merged without human review

Executive summary

AI now writes a large and fast-growing share of the world's code. This report asks whether organizations control what the machine writes. It is based on 112 million commits from 340,000 repositories in 35,203 organizations in 159 countries, analyzed with the AgenticReview method, which condenses 34 measures into 8 indicators, four dimensions and one composite index for agentic code control (CIACC, 0-10).

Usage has exploded. In one year, the share of organizations with agent commits in a given month nearly quadrupled, from 9% to 34%, and the agents' share of all commits rose from 4% to 21%. 69% of the organizations studied show traces of AI agents in their code work. Usage rises with size, from 50% among one- and two-person teams to 94% among organizations with 50+ developers, and is highest among AI and SaaS companies and in North America and Oceania.

Control varies widely. The median organization scores 6.4 out of 10 on the CIACC. The top tenth (Code with Control) reaches 8.3 or more, the bottom tenth (Code without Control) 4.0 or lower. Size, industry and continent all matter, with spans of 1.3, 1.0 and 1.0 points. Western AI and SaaS companies lead, while six of the eight weakest combinations of continent, industry and size are Asian.

The framework exists - human control does not. 69% of assessable organizations have a strong governance framework around their agents, but only 41% reach a high level in security, 40% in dependency management and 32% in architectural order. In a line-by-line sample, 75% of the pull requests opened by agents were merged without human review, 98% of the agents' lines were left untouched within a week, and only 1.2% of the untouched lines were fixed within three months.

Control cannot be inherited. Size buys routines - tests, security and governance - but not architecture, and it costs speed. Strength in what classic code review measures explains only 7% of the variation in the agentic dimensions: control in the AI era has to be measured and built in its own right.

Coding is solved. Its control is not.

Cite this report

The State of AI Code Control 2026 (the CiSCiN report). AgenticReview benchmark, November 2026 edition (model v4.10, thresholds 2026-11a). Author: AgenticReview Group. Licensed under CC BY 4.0. DOI: 10.5281/zenodo.23966831