The new Implementor Agent can then take that technical plan and write the code for it.

"The developer can choose an agent to implement the code that it will come back with before a pull request is raised," Devenny said. The dev agent notifies the developer, shares the diff, and "the developer gets to review it and make any changes." This often involves a conversation between the developer and the agent.

Use Cases for Custom Dev Agents

While there are those 20 or so out-of-the-box AI agents, Valliani said early adopters demanded a blank canvas to build their own, each dedicated to a singular purpose.

With that in mind, Atlassian added the no-code ability to build your own agent, which even nontechnical departments at Atlassian have used. Internally, Atlassian created its own Onboarding Buddy agent that, within the first month, was answering 2,000 questions from 70% of Atlassian new hires.

One Atlassian engineering team built and then internally shared its own Pull Request Generator that team members said automated 80% of developer tasks — like adding feature flags and test automation — cutting the average pull request down to 30 minutes.

This ability to create custom dev agents also allows customers to enforce organizational and industry standards, often pulling from existing Confluence documentation.

"Then you can tailor these agents, can specify knowledge sources and constrain the agent, then run agents automatically," Valliani said.

Facing a time where more AI-generated code means more demand for code reviews, a third early agentic AI developer use case is addressed internally at Atlassian with its Code Reviewer, which combines a pull request with a Jira issue, Devenny said, to "deduce the acceptance criteria in the Jira issue, and we show whether the code actually matches that acceptance criteria or not."

Code Reviewer also checks for syntax errors and typos. If it’s a code change to a more sensitive or crucial piece of the codebase, it can be flagged for review by two reviewers.

Last week, Code Reviewer suggested Devenny reduce the complexity of this function and provided a code example of how to do it. With this in mind, in the coming months, Bitbucket will be introducing an "apply suggestion" button.

Since the adoption of this particular dev agent in late October, Atlassian engineering has experienced a 45% improvement in PR cycle time for any pull request when Code Reviewer is being used, versus when it is not.

"The reason it’s such an aggressive jump decrease," Devenny said, "is because you don’t have to wait for hours or days for a peer to jump in, you immediately have feedback."

"At the moment, the tasks that we’re doing for developers are like repetitive tasks, ones that developers are like not enthused about doing in the first place. Being able to remove that from their things that they have to do means that they get to spend more time focused on the tasks that they love implementing."

— Josh Devenny, Atlassian

Ops Guide is another dev agent that helps out with on-call duties, seeking to reduce time to detect, respond to and recover from incidents.

Each of these dev agents helps decrease overall issue cycle time, but does not automate out the essential human in the loop.

Devenny also mentioned that Pipeline Fixer and Deployment Summarizer dev agents are coming soon.

The Feedback Decoder AI agent creates and analyzes transcripts from customer feedback from the Loom video messaging platform and other sources, which then can be used to automatically create Jira tickets. The Jira Product Discovery (JPD) AI agent can then connect Jira tickets with customer data, which are in turn tied to stated Goals. The Feedback Decoder can also update teams in Slack, log activity in HubSpot and more.

"JPD makes it really simple to see how your product idea moves the needle on important goals and metrics, so you can explain how your great product idea impacts your business, and you get to decide who has access to your own maps," said AI native approach to software development really changes the game. It breaks down traditional silos between roles and gets rid of repetitive tasks," Bharadwaj said. "Imagine a world where every team member can do more than their traditionally assigned role. Powered by our Rovo agents, this creates a team of powerful generalists ready for an AI native world."

Studio also allows teams to implement multiple issues at the same time, such as "generate unit tests for these four things," and the agent will create four implementations, one for each issue, and move it to review, or even set up an automation.

Recently, Atlassian engineering had to update an API across 200 services. A custom dev agent raised 200 pull requests in one hour.

"People have ‘aha’ moments or ‘wow’ moments where they’re like, ‘Oh, I won’t need to do this type of task anymore.’ And that’s a huge mind shift. I have seen it in a positive light in every single case, because, as the capability of LLMs increases, we will be able to do more and more tasks," Devenny said.

"At the moment, the tasks that we’re doing for developers are like repetitive tasks, ones that developers are not enthused about doing in the first place. Being able to remove that from their things that they have to do means that they get to spend more time focused on the tasks that they love implementing."

Rovo dev agents are now in open beta for Atlassian customers on GitHub and Bitbucket. While in beta, there is no cost to the customer.

Disclosure: Atlassian paid airfare and lodging for the reporter to attend this conference.