Key Takeaways

  • Atlassian compressed standard six-month enterprise feature timelines down to six weeks for Remix with Rovo and eight weeks for Confluence Slides.
  • A non-technical product manager named Ya checked in 26 pull requests in a single month after engineers set up an isolated front-end development environment for her.
  • The team integrated Arize for prompt debugging, doubling evaluation throughput across their large language model workflows.
  • Design QA was automated by connecting Figma MCP with coding agents, fixing visual mismatches directly in code without manual bug filing.

The Method

Enterprise feature work usually drags because front-end adjustments create friction between product, design, and engineering. Atlassian Chief Product Officer Tamar Yehoshua shared how the Confluence team dismantled this bottleneck by changing who writes code and how design QA happens.

First, engineering gave product managers direct access to front-end changes without risking production infrastructure. “This PM Ya had never written any code, had never even used a terminal before, but she sat down with the engineering partner who actually built a harness for her to use that made it easier for her to check in code in the front end,” Yehoshua explained. By removing the fear of breaking the build, the PM took over UI copy, styling tweaks, and layout adjustments. “She ended up checking in 26 PRs in a month, which was more than most of the engineers on the team.”

Second, the team stripped manual prompt debugging out of engineering sprint cycles. They used Arize to run prompt evals and trace model responses, resulting in a 2x throughput boost. This let engineers test model changes against concrete evals instead of running ad-hoc manual tests.

Third, they automated the visual QA handoff between design and code. Instead of designers taking screenshots, opening tickets, and waiting two sprints for visual fixes, the team connected Figma to their repository via Figma MCP. “They automated the fixing of design bugs. How many times have you seen the code and it didn't match the design? So they used Figma MCP and mapped the designs from Figma to the actual code and then automated the fixing with a coding agent,” said Yehoshua.

Combining these three workflows produced massive speed gains. “What would have taken before AI about 6 months now took 6 weeks with much tighter iteration loops,” Yehoshua noted.

Where This Breaks Down

This system falls apart if engineers fail to isolate the front-end environment properly. If a PM commits unvetted logic directly to core production code, code reviews become a slog. Engineers end up spending more time fixing messy pull requests than they would have spent building the UI themselves.

Automated design agents also struggle when designs lack strict component consistency in Figma. If a design team uses detached components or inconsistent auto-layouts, an agent will misinterpret the layout and produce bad CSS. The automation only works when your design system is as structured as your code.

What to Do With This

Pick one front-end repo your team touches daily. Have an engineer create a sandboxed local environment and a git workflow template specifically for a non-technical PM this Friday. Start with UI copy and spacing adjustments, and require one small pull request from the PM next week.