Key Takeaways

  • Allowing non-engineers to push code directly into a 20-year-old codebase creates stability risks and compliance failures.
  • Atlassian replaced manual spec writing with Loom screen recordings, automatically turning Figma walkthroughs and voice tracks into structured Jira work items.
  • Cloud-managed coding agents used these items to write code against standard design systems, while Slack bots routed live bugs straight to agents for fixes.
  • A dedicated robo-agent sorted and categorized over 900 pieces of customer feedback without manual human filtering.
  • The workflow allowed the team to ship 22 user-facing features in 10 weeks, tripling their historical throughput from in-progress to shipped.

The Method

Every tech company wants faster release cycles, but enterprise software has a safety problem. “Jira has been around for over 20 years. It's a huge complicated codebase,” says Tamar Yehoshua, Chief Product Officer at Atlassian. “So this instance of PM checking in code can be kind of dangerous. So we didn't want to do that.”

Instead of letting product managers write raw code in production, Atlassian automated the loop between product intent and agent execution. The workflow operates in four clear stages.

First, product managers skip the traditional product requirements document. They record a Loom video showing the user interface, walking through Figma designs with a voiceover, or talking through a whiteboard session. “First they started with Loom,” Yehoshua explains. “And in a Loom video, you could record the UI that you wanted to change or you could record here the Figma designs with a talk track or you could just brainstorm that goes into a recording and then Loom has the ability from the recording to automatically create work items.”

Second, those auto-generated work items trigger cloud-managed coding agents. The agents do not touch raw production repositories without guardrails. They build prototypes constrained by Atlassian's established design system components, keeping the interface consistent and compliant.

Third, internal testing feedback gets routed directly into the loop. “We took the feedback and the bugs that came in through Slack and we used the Jira agent in Slack to triage them and then automatically send them to the coding agent for fixes,” Yehoshua notes. When a tester posts a bug in Slack, the Jira bot identifies the problem, drafts a repair ticket, and commands the coding agent to resolve it.

Fourth, customer feedback loops run at scale. “We got over 900 pieces of feedback,” Yehoshua says. “We used a robo agent to then triage and categorize them. And we were very happily surprised with the quality that we got of the triage.” The result was clear: “And from in progress to shipped their throughput was about 3x what it normally was and they shipped in about 10 weeks 22 userfacing features.”

Where This Breaks Down

This pipeline only works when your design system is strictly enforced. If your components are poorly documented, inconsistent, or scattered across disparate repositories, coding agents will invent their own CSS and custom logic. That will pollute your codebase faster than human engineers can clean it up.

It also requires tight scope control. Agents excel at modular UI changes, form adjustments, and workflow optimizations where inputs and outputs are clear. They struggle with deep architectural modifications, multi-service database transactions, and security-critical authentication logic. If you feed an architectural rewrite into an agent via a screen recording, you will get broken dependencies and silent regressions.

What to Do With This

Stop writing five-page requirement documents for small UI improvements. Tomorrow morning, record a two-minute video walking through a specific screen flow you want improved, describe the desired changes aloud, and use an automated transcription tool to generate the implementation steps. Hand those discrete component updates to an engineer or coding tool with strict instructions to use your existing design tokens.