On September 22, Atlassian ran a webinar called State of AI: SDLC. It's been over a month that I've convinced myself of writing this article to share with the community, and finally got to it.
Atlassian called it the largest simultaneous digital event in the company's history, with more than 30,000 registrations. Mike Cannon-Brookes opened with Vercel CEO Guillermo Rauch, and breakout sessions with DX, Dropbox, Lovable and 1Password followed.
A colleague and I attended separately and compared notes afterwards. It was interesting to share what caught our attention. Below are the five ideas that stuck with both of us, and at the end there are a few questions I'd like this community's take on.
1. Adoption is basically solved. Impact isn't.
These numbers set up the rest of the day:
- 94% of organizations use a coding agent for at least some tasks
- 6% have something close to an AI-native development lifecycle
- 90%+ of pull requests across DX's customer base are now AI-assisted
- …yet most teams report only 10–20% efficiency gains
Atlassian's own breakout offered one explanation: writing code is only about 16% of engineering work. If you make that 16% faster, you still have to deal with planning, handoffs, review and waiting on CI. The Dropbox/DX session called it "death by a thousand cuts."
There was one more stat that really caught my attention. Engineers self-report saving around 6 hours a week, but perceived productivity runs about 40% ahead of measured output. I felt that.
2. The bottleneck is context, not the model
Nearly every session came back to this point: an agent can only be as good as the information it's given. The DX talk "Your agent is only as good as your context" offered a checklist called CAFE(S):
- Clarity: can the agent understand the request without guessing?
- Actionability: does it know what "done" looks like?
- Fidelity: is the context accurate and current?
- Efficiency: is it lean, or is the useful part buried?
- (Security): should the agent have seen this at all?
The failure mode that hit closest to home was specification ambiguity. If you hand an agent a fuzzy ticket, it quietly picks one interpretation, and that early assumption snowballs. We've always had the vague-Jira-ticket problem. Agents just make it expensive faster.
Atlassian shared that grounding its own agents in Teamwork Graph improved answer quality by 44% and cut token use by 48%. The framing I liked best: engineers can use Cursor, Claude, Copilot or whatever suits them, because the agent is interchangeable, but the work item is not.
3. "Slow down" is a new kind of customer feedback
This one surprised me most. In the main session, Atlassian said several customers have asked it to ship more slowly. They said that hadn't happened in about 22 years in the industry.
When building is cheap, the hard part becomes deciding what deserves to exist. The product-decisions panel said it plainly: PMs are the new bottleneck. The questions they suggested asking were "Why does this exist? Why does it exist here?" They also recommended iterating on one feature 2–4 times rather than shipping more features.
The DX data backs this up. The innovation ratio (the share of work that delivers new value) rose only about 1%, even though AI spend at larger companies is up roughly 28x and throughput is up about 25%. Most engineering leaders are using the extra capacity to pay down tech debt and add quality gates first.
4. The pressure moved to code review
Some of the numbers from Dropbox and DX:
- Median PR size roughly doubled, from about 40 lines to 70+
- Maintainability went up but change confidence went down: engineers can follow the code, but they don't fully trust it
- Incremental delivery had the biggest single drop on DX's developer experience index
- Cross-team collaboration fell, partly because engineers now talk to agents instead of to each other
Dropbox's fixes were mostly human ones: more upfront architecture reviews by senior engineers, scheduled design sprints, and protected focus time. Atlassian reported a 45% cut in PR cycle time from an AI reviewer that checks PRs against codified standards. 1Password's point was blunt: markdown files full of style suggestions don't work for agents, but linters that pass or fail do.
5. If an agent wrote the code, who gets paged?
The panel said that's the wrong question. You page the team that owns the service, whoever or whatever wrote the code. The better question is whether the agent had enough context about the service and its recent changes.
Here's a practical test I'm planning to borrow: Can you answer "what changed on this service in the last 2 hours, and who or what changed it?" in under 60 seconds?
And the line I wrote down twice: "Agents don't create ownership problems; they make pre-existing mess visible at machine speed."
My takeaway
AI didn't change what good teams need (clear work items, owned services, small batches, honest feedback loops). It raised the cost of not having those need served.
What's your experience so far with the new 'AI SDLC'?
What are your org's major challenges for adoption? and for scaling?
Where did your AI time savings actually go? Did you get more features, less tech debt, or did the time just disappear?
Has AI changed how your team writes Jira work items? Are definitions of done getting tighter now that an agent reads them literally?
Have your PRs grown? If so, what has actually helped your reviewers keep up?
Share your answers!
