In internal evaluations, Atlassian found that code mode in the Rovo agent harness reduced the latency of complex Jira-related queries by over 50%, cut token consumption by 55%, and improved accuracy by 30%. These metrics were measured against a baseline harness that used standard meta-tools instead of code execution across a representative set of complex queries.
Rovo Chat began as a tool for searching internal Atlassian data and was expanded to retrieve information from external systems; it now supports over 50 third-party data sources. Atlassian wrote that "Every user should have access to an always-on sidekick capable of tackling large tasks, planning work, and managing complex, cross-cutting workflows." To power that shift it rethought its execution model, moving beyond simple synchronous tool calls into a distributed environment that combines a split-plane architecture, isolated compute sandboxes, asynchronous multi-agent coordination and automated self-evolution.
Atlassian separated the data plane and the control plane so that each can scale independently and degrade gracefully. The user connects directly to the control plane, so if a sandbox crashes the agent retries and spins up a new sandbox container transparently. For simple operations a smaller sandbox instance is used and upgraded if the required compute increases, and the control plane can start a larger sandbox for heavy tasks. The post says this design means one scaled service will handle events for millions of users.
Because relying on an LLM to invoke tools sequentially and handle pagination over large datasets of Jira work items, Atlas goals and external data is inefficient, the harness offloads data iteration and pagination to code execution in the sandbox. Combining programmatic tool calling with progressive disclosure means Atlassian can safely expose thousands of API actions to the agent across first-party and third-party products. A CI/CD pipeline versions and packages tools into a Python package that updates the sandbox on a warm start, maintaining sub-100ms startup times, and the sandbox has programmatic access to both statically defined internal tools and MCP servers managed by third parties. A callback bridge routes every tool call through Atlassian's internal server, where user permissions are validated and requests authenticated before any external API is invoked.
Interactive sub-agents let a parent agent delegate work to background processes; a sub-agent can outlive the parent's response turn, keep its own conversation history, and wake the parent when it finishes. For error recovery the harness uses discovery probes that return structured feedback, and checkpointing that snapshots execution state so failed multi-step automations can be rolled back and retried. Failed queries are flagged and fed into the evaluation suite, with engineers reviewing and approving optimizations before they are applied.