Atlassian is formalizing an AI-native SDLC pattern: a shared context graph, explicit agent scopes, automated standards, and an AI reviewer wired directly into pull requests. Agent loops in Jira watch for structured work, delegate to a coding agent, and bring back tested changes for human approval. Because the same layer that gives context also defines what agents can touch, policy and capability stay aligned. This isn’t another chat interface; it is workflow-native automation with a merge gate and observable impact.

Four primitives determine whether agent loops scale safely: contextual grounding, fine-grained access, consistent standards, and independent review. Code Context anchors agent behavior in real architecture and decisions. Agent Context Controls scope visibility and write permissions per space. Standards encode code quality, security, and design norms and map them to repositories. AI Review acts as a consistent first-pass reviewer, catching violations before human time is spent. Together, these controls let platform teams raise or relax fidelity and risk tolerance without re-plumbing the stack.

Assign clear ownership. Platform teams own context curation, standards baselines, and access policies. Teams opt in with loop-ready issues: crisp intent, scope boundaries, and test expectations. Define SLOs for time-to-PR, first-pass acceptance rate, and blast radius (e.g., files or services touched per change). Schedule change windows for updating standards or context to avoid churn. Maintain golden repos and seed tests to benchmark quality drift. Finally, wire loop events to incident and audit pipelines so exceptions, rollbacks, and cost anomalies are detectable within hours—not quarters.

Common pitfalls include agents wandering beyond intended scope, context drift after structural refactors, and PRs that look compliant but degrade runtime behavior. Mitigations: tighten Agent Context Controls to explicit repo and path scopes; gate merges on policy plus targeted runtime checks; rotate context refresh jobs after dependency or interface changes; and apply sampling-based human review on risky categories (auth, billing, infra). Track cost-per-accepted-PR and mean rework to catch low-quality loops early. If a loop’s acceptance rate dips, freeze it, retrain standards on fresh exemplars, and relaunch with narrower scope.

Day 0–14: Select two services with clear tests and low coupling. Map Standards, seed AI Review with policy examples, and bound Agent Context Controls to target repos. Establish baselines: time-to-PR, review load, defects, and cost. Day 15–45: Enable agent loops for backlog chores (lint, config, small refactors), track acceptance rate, and tune standards. Day 46–90: Expand to medium tasks with integration tests, introduce a change window for policy updates, and present results via usage and impact dashboards: throughput, rework, defect escape rate, and cost-per-merged change.