AI NewsProductivityAnnouncement

Atlassian says AI now writes 69.7 percent of merged code

Atlassian and DX published new figures on AI code at 500 firms. 340,000 engineers are in the panel. AI wrote 69.7 percent of merged code in the latest month.

AI News

Editorial3 min read

LinkedInX
Atlassian AI SDLC Index hero graphic

Image: Atlassian

Why it mattersThe old story was AI writes code cheaply. The new figure says the cost to merge a unit of code tripled this year.

Atlassian released a company data set on Tuesday. It measures how AI has changed software work at more than 500 engineering firms. The set is called the AI SDLC Index. It uses system data and developer surveys. The data comes from the DX engineering intelligence platform. Atlassian gathered it between August and October. The firms employ more than 340,000 engineers in total.

The headline figure is for AI-authored code. It rose from 49 percent of merged code in April to 69.7 percent in October. These are Atlassian's own numbers against its own panel. Code survivorship is a second figure. It is code that survives past the first commit. It dipped to 59 percent in September and returned to 68.2 percent in October.

The cost moved in the other direction

A second line in the data shows the cost per unit of merged output. It rose from $8.85 in April to $29.11 in October. The measure is AI spend per useful unit of code. Atlassian says the October figure is 28 percent lower than September. The September peak was about $40. The author adds that this measure is early. It will move as better tools arrive.

The index also reports time lost to friction. Developers lose about 20 percent of their work year to it. Atlassian puts the figure at 409 hours per engineer. Two large groups inside that number are equal. Review wait time is 86.5 hours. Build wait time is 86.5 hours. Time on new features has stayed at about 58 percent since April.

Review load grew with the code

Monthly review load per developer moved too. It went from 12.8 pull requests in January to 22.3 in October. Median time to first review runs on a scale. It is 31.2 minutes at firms where AI writes 20 to 40 percent of code. It rises to 97.8 minutes where AI writes more than 80 percent. In October 21.2 percent of all pull requests got a bot review. At the start of the year the figure was in the single digits.

Favorability scores from the survey moved on two items. Documentation satisfaction rose from 47.7 percent to 60.4 percent year on year. Production debugging rose from 52.9 percent to 61.6 percent. CEO Mike Cannon-Brookes writes that "usage and spend are easy to count, but they're not outcomes."

The reproducible part

Atlassian says the next index will publish next quarter. The method worth copying is the split between system data and survey. DX pulls data on commits, pull requests, review wait time and build time. The survey asks the same engineers about their own experience. Both together are harder to game than either one alone. A team can set the same tools on its own code. It can then run both sides for a quarter and get its own figure.

Three numbers are worth watching on the next release. First, whether AI-authored code share keeps rising. The point to watch is about 70 percent. At that share every repository already has mixed authorship. Second, whether code survivorship stays above 68 percent. Harder tasks are being handed to agents, so the figure may fall again. Third, whether review load keeps climbing. Review has already become the slowest step in the panel. A team that merges agent code without a review plan has a slow step coming.

Source

SourceAtlassian

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project