Why better tickets help agents write better code
Atlassian’s internal experiment found AI-native Jira tickets improved engineering output fivefold, with higher quality and agent-readiness compared to traditional tickets.
Atlassian’s team built an enterprise compensation planning tool using AI agents, aiming to test how quickly a small group could deliver a production-grade product. They abandoned traditional sprint structures and instead relied on detailed Jira tickets to guide autonomous agents. The approach yielded roughly five times more output per engineer than a comparable traditionally built product, measured across code, complexity, and integrations. The team also authored nearly five times more Jira tickets per person during the same period, reflecting a shift in workflow intensity and granularity.
The quality of AI-native tickets significantly exceeded traditional ones, scoring 4.47 out of 5 in a blinded evaluation compared to 2.72 for traditional tickets. Eighty-three percent of AI-native tickets were deemed ‘agent-ready,’ meaning they contained explicit acceptance criteria, scope boundaries, and contextual details necessary for autonomous execution. The widest quality gaps appeared in areas critical for agents: acceptance criteria and context, which were often absent or vague in traditional tickets that typically consisted of a single-line summary and a link.
Agents frequently uncovered latent issues such as bugs or technical debt during execution, but without a way to capture these findings in real time, the insights were lost when sessions ended. This led to recurring rediscovery of the same problems, increasing operational friction. The team is now addressing this by integrating agent activity directly into Jira tickets, allowing agents to log technical debt or findings immediately rather than losing them upon session closure.
The experiment demonstrated that high-quality, explicit Jira tickets became the backbone of Atlassian’s AI-native workflow, enabling faster scoping, coordination, and handoffs without formal mandates. Enterprises already entrust Atlassian with sensitive planning data, and maintaining this trust requires keeping agent artifacts, context, and decisions within the same controlled environment. The findings underscore that the specificity of instructions in tickets directly shapes the quality of outcomes, as every assumption left for agents to interpret incurs downstream costs.