Software program groups have taken to coding brokers quicker than virtually any software earlier than them, and but delivery software program hasn’t sped as much as match. Atlassian has put a quantity on the hole. In a longitudinal study the corporate ran with the engineering intelligence platform DX, the quantity of AI-authored code practically doubled over three months, whereas developer productiveness positive aspects stalled – topping out at round 15%, and coming in beneath 10% at many organizations.
On the again of this comes an announcement from Atlassian with a set of new capabilities in Jira and the Teamwork Graph aimed on the work that surrounds the code fairly than the coding itself. Ming Wu, Head of Engineering for Dev AI at Atlassian and joined round two years in the past from Microsoft and GitHub, walked me by way of the reasoning forward of the launch. Her start line is that writing code was by no means the place a lot of the effort went.
The coding half, it is a necessary half, however it’s solely fifteen to sixteen p.c of the time. All the remaining comes from planning and iterating on the design and alignment, and addressing all the problems when it goes out, and the code assessment.
The remainder of the software program improvement lifecycle is the place the fee lives, and Wu says that is the half {that a} lone coding agent tends to deal with badly.
The place the human holds the wheel
Atlassian’s shorthand for the shift is that work turns into human-steered and agent-executed. Requested what steering means whenever you take the phrase actually, Wu was clear that the senior engineer’s job does not shrink.
I truly do not suppose the human’s position, the primary position, has modified. The human is the one holding the duty of how this factor will ship, and what you are going to ship, or when. AI instruments, it’s extremely quick, however it does not know what to do.
What modifications is the typing and the small, well-defined duties round it. She explains:
I haven’t got to sort anymore, however I nonetheless must see. I’ve to inform it what to do.
That leaves the repetitive work – tech debt, feature-flag clean-up, small atomic fixes – because the half an agent can take up, and frees individuals for the extra complicated and inventive work. The purpose she emphasizes is accountability – a human nonetheless has to log off on no matter ships.
Atlassian says that for this shift to truly work, three issues must be true. Intent must be structured earlier than work begins, so an agent will get the requirement, the structure, the choice historical past and the constraints, fairly than a one-line abstract. The agent ought to be a runtime alternative fairly than a course of one, so a workforce can attain for the Cursor built-in improvement atmosphere on one process and Claude Code on one other with out the workflow forking every time. And autonomy has to remain observable, so agent work does not vanish into terminals and browser tabs.
Why context sits on the middle
Working by way of the announcement is the Teamwork Graph, which Atlassian describes as its context layer – a map of labor, code, individuals, choices and dependencies that pulls collectively the duty in Jira, the necessities in Confluence, the dialog in Slack, the code context from GitHub, and buyer perception from Jira Product Discovery. Wu observes that the context layer is more durable to construct than it sounds, and she or he believes it issues greater than the brokers themselves. She elaborates:
The context layer shouldn’t be concerning the uncooked knowledge. How do you truly retrieve, effectively and neatly, solely the related ones? Everyone knows paperwork go out of date.
She calls the work of choosing the context that is truly related as context engineering, and she or he’s put a great deal of Atlassian’s funding there. She was additionally express that it is a declare many corporations make and few comply with by way of on, and she or he leaned on her personal background to say so:
I come from Microsoft, and I used to be in GitHub. The context graph is all the time the laborious subject that each firm claims they do. However not everybody truly places in critical effort.
With out it, Wu notes:
An agent will not work effectively in case you simply throw all the things collectively.
Starved of context, based on Atlassian, an agent solves a ticket too actually, misses an architectural constraint, or produces a pull request that appears believable till a senior engineer spends an hour unwinding it. Jim Mercer, Program Vice President for software program improvement, DevOps and DevSecOps on the analyst agency IDC, made the identical level:
Brokers working with no deep understanding of workforce choices, architectural constraints, and mission historical past produce misaligned code extra rapidly, resulting in technical debt and manufacturing points.
Stepping into the main points, these are the weather which might be delivery:
Getting the context in. Jira Planner brings spec-driven improvement into Jira. For complicated initiatives, Atlassian says it pulls from the codebase, from Jira and Confluence historical past, and from workforce context to provide a structured technical specification in Confluence, readable by an individual and usable by an agent. Wu describes the goal as brownfield work, the tangled current programs the place an autonomous agent tends to come back unstuck, and she or he does not faux the individuals might be designed out of it, explaining:
Particularly from what I see in banking and finance, that sort of extra rigorous atmosphere with a whole lot of dependencies, you can not skip the complexity of the real-world downside, and likewise the collaboration. You truly do want the individuals coming in.
Alongside it, Jira for Slack turns a thread into a piece merchandise when a workforce asks @Jira, retaining the dialog that formed a choice hooked up to it. Loom video prompts take a display screen recording and a spoken walk-through, and switch them into structured directions an agent can act on.
Delegating and watching the work. Brokers in Jira lets groups assign a piece merchandise to Claude Code, Cursor, or GitHub Copilot instantly, with OpenAI’s Codex listed as coming quickly. The Jira Coding Agent, now constructed into each paid Jira plan at no additional price, takes a well-scoped merchandise and returns a ready-to-review pull request for routine fixes with no developer dropping into a neighborhood atmosphere. Agent classes in Jira, along with new Teamwork Graph command-line interface hooks, hyperlink a neighborhood terminal session again to the Jira merchandise it belongs to, so the file survives a closed laptop computer.
Governing the fee and the size. Coding agent automations let groups route routine work – bug fixes, vulnerability remediation, check technology, documentation updates – by way of Jira’s automation rule builder, with an engineer notified when a pull request is prepared. The Agentic Engineering mission template and a setup wizard rise up an agent-ready mission with workflows and integrations pre-configured. DX AI price administration gathers spend and token knowledge throughout instruments similar to Claude, Cursor and GitHub Copilot alongside Jira, and estimates a price per pull request.
The numbers, and what they measure
Atlassian ran these patterns throughout its personal engineering group earlier than delivery them. In an inside research of 6,000 engineers utilizing the brand new capabilities, the corporate reported a 44% rise in agent process completion effectivity, a 48% drop in token consumption, a 36% discount in pull request cycle time, and 51% of routine code vulnerabilities resolved autonomously and queued for developer assessment.
The work is finished by an agent, after which it waits for an individual to assessment it earlier than it merges. For atomic, well-scoped duties similar to vulnerability tickets, Wu places the interior automation price within the area of half to two-thirds, including:
The extra complicated the duty is, you’ll have to work together and inform the agent.
We additionally talked about benchmarks and the way usually a ready-to-review pull request merges with solely mild assessment. Wu factors to the SWE-bench coding benchmark. She notes that the complete set sits within the low 40s and the verified set far increased, however does not put a lot weight on it, observing:
Truthfully, I feel SWE-bench is saturated. It is also open, so everybody sees that. Take the numbers with a grain of salt. We’re not banking on that an excessive amount of.
On technique, she locations Jira because the layer that sits throughout these brokers fairly than a rival to the mannequin builders, and agrees that a number of the positioning remains to be transferring.
Jira is effectively positioned to be that orchestration layer. It must be, as a result of no person else can be in that position. Atlassian shouldn’t be a mannequin participant. I do not suppose that is our competing technique proper now.
She ends on the purpose that the collaboration solely works as soon as an agent might be trusted with the context a colleague would have:
You’re employed together with your teammate solely after they perceive your context. Then you’ll be able to belief them to truly work along with you.
Atlassian says Brokers in Jira for Claude Code, Cursor and Copilot, together with Jira for Slack, the Jira Coding Agent, agent automations, the agentic templates, and agent classes, can be found now for paid Jira Cloud prospects at no further price. Jira Planner is in early entry, and Codex in Jira is listed as coming quickly. DX AI price administration is offered to Atlassian DX prospects.
My take
The agent clears the routine vulnerability, and an individual nonetheless seems to be earlier than it merges. Utilizing a metric of “queued for developer assessment” looks like a a lot easier rationalization of human-steered and avoids hype.
Earlier this 12 months I talked to a number of Atlassian consultants concerning the Teamwork Graph pitch and spent a very good a part of these conversations pushing on whether or not it was actual but or nonetheless a keynote slide – how recent the context truly was, how briskly the graph knew when one thing had modified. Just a few months on, each agent functionality on this launch leans on it, and the context layer is now carrying the entire product.
Atlassian has put its coding agent in each paid plan, and its actual effort into context and governance, fairly than becoming a member of the continuing vendor agent race. The reminiscence of what your workforce determined, and the file of who signed off earlier than it shipped, are Atlassian’s strong floor, due to its sturdy historical past because the system of file.