Agents touch 27.7% of code changes at the median company

On 27 October 2026, Jellyfish released a report on engineering work. Jellyfish is a Boston company that collects data from tools such as Git and Jira. It found that in September, AI agents opened or committed code to 27.7% of pull requests, which are proposed code changes, at the median company. In January it was about 2%.

How it works

A software factory is a pipeline where agents select, write, check and ship the work. In most pipeline systems, a person continues to select the work. Jellyfish says that teams that change the instructions of their agents more frequently merge more changes.

The numbers

The data includes more than 1,300 companies.

• AI tools wrote 29% of code at the median company.
• Software factories opened 9.4% of pull requests at the 99th-percentile company.
• AI alone checked only 0.5% of merged pull requests.

What is not known

In a poll of developers, 88% said that AI tools for code help their work. Only 62% said that their employer controls the risks of AI code. Jellyfish says that only a small number of companies can use agents in all teams.

Insight: Checks by a person hold the line, but risky changes can skip them

Jellyfish finds that AI alone checked 0.5% of merged changes. Different research found that 43.1% to 56.2% of agent changes to code and deployment systems together merged with no check. Mintlify says that 9% of companies let agents publish text with no check by a person. A check by a person is the last control, and it seems weak on risky changes. Select which changes can go without it.