44% more accurate on fewer tokens, and issues that become pull requests
Atlassian's coding agents for Jira report better accuracy at lower token cost and convert issues directly into pull requests. Closing the loop inside the tracker is the part that changes behaviour.
Atlassian's internal benchmarking reports that agents enriched by its Teamwork Graph produced 44% more accurate results while using 48% fewer tokens than agents without that context. Jira work items can now be assigned to a coding agent that opens a draft pull request.
More accurate and 48% cheaper is worth pausing on, because those usually trade against each other — accuracy is normally bought with more reasoning tokens. The stated cause is the context layer, not the model: giving the agent the right subset of the codebase and the ticket history rather than more of everything. That is the whole argument for retrieval over raw context length, with a number attached.
The issue-to-pull-request path is the more consequential feature. Coding agents have mostly lived in the editor, which means a developer must already be working on something for the agent to help. Starting from the tracker means work can begin from the place work is actually assigned, without a human opening a branch first.
That changes who can start a task, and it is where the caution belongs. A pull request generated from a ticket carries the ticket's ambiguity into code, and tickets are ambiguous — they are written by people who assume a reader with context. The review burden shifts to whoever merges, and there will be more to review.
Both figures are vendor internal benchmarking against their own un-enriched baseline, which is the standard for this category of claim and should be treated accordingly. The workflow change is real regardless of whether the number survives contact with someone else's codebase.
The New Stack — Atlassian wants developers to finally like Jira → · DevOps.com — Atlassian Extends AI Reach of Jira Into Agentic Engineering Workflows →