What High-Throughput Engineers do Differently and Why AI Widens the Gap
A study of 15 high-throughput engineers at Atlassian found AI amplifies existing best practices rather than replacing them, widening the performance gap between disciplined and ad-hoc teams.
A study of 15 high-throughput Atlassian engineers working on complex legacy systems found that AI did not replace core engineering fundamentals but accelerated their effectiveness. Engineers who consistently scoped work into small, single-responsibility changes saw faster builds, fewer merge conflicts, and improved reliability when using AI agents. The shift toward smaller, atomic pull requests reduced cognitive load and made changes easier for both human reviewers and AI systems to evaluate and test independently.
Structured intent, traditionally documented for human review, now serves as executable context for AI agents. Engineers who formalized specifications—mapping requirements, use cases, and domain logic—achieved more reliable AI outputs. Newer frameworks like BMAD repurpose this discipline for agent-driven workflows, while graph engineering formalizes explicit workflows to guide AI execution. Teams that decomposed projects into tightly scoped Jira work items and assigned them to parallel coding agents reported reduced overlap and fewer conflicts, with human reviewers focusing on integration and quality gates.
High test coverage and realistic mocks emerged as critical enablers, enabling AI to autonomously test and fix errors while reducing manual review burden. Engineers who rebuilt test infrastructure to mirror production behavior achieved 90% test coverage from Storybook, transforming coverage into a living specification for AI agents. Fast, reliable build and test pipelines were cited as more impactful than specific AI tools, with automation handling routine tasks like stale feature flags and dependency updates to protect deep-focus work.
Parallel work management required disciplined context switching, with engineers using tools like multiple git worktrees or remote dev environments to maintain warm workflows without losing state. Leadership investment in high-quality remote access and flexible tooling was identified as essential, particularly for engineers on slower hardware. Review velocity became a bottleneck as AI accelerated code generation, with reciprocal, domain-routed review networks and team-level service-level objectives ensuring fast, high-quality feedback loops remained a shared capability rather than an individual burden.