Atlassian's internal code-review evaluation

Atlassian deployed Rovo Dev Code Reviewer internally as a human-in-the-loop pull-request reviewer. Its April 2026 engineering account says the system used Anthropic Claude 3.5 Sonnet to generate review suggestions, alongside separate quality checks. The account reports an online evaluation spanning more than 1,900 repositories over roughly a year.

Atlassian reported a 30.8% reduction in median pull-request cycle time with its Claude-backed Rovo Dev Code Reviewer and 35.6% fewer human-written review comments. It also says 38.7% of generated comments led to a code change in the next commit, versus 44.45% for human comments. These are product-level results from Atlassian's own observational evaluation, not a controlled estimate of Claude alone's causal effect or of Rovo Chat's Long Horizon feature. The work involved a Monash University collaborator and was accepted at ICSE 2026, but that does not make the company-run deployment metrics an independent audit.

Related

• atlassian-rovo-claude-long-horizon — a distinct Rovo Chat capability using Claude-class models.
• atlassian-claude-rovo-agent-platform — Anthropic's account of Atlassian's broader multimodel platform.
• 1password-codex-engineering-rollout — a different engineering productivity case.