Rovo Chat has expanded from an internal data search tool to a distributed agent harness deployed at scale, processing requests for what Atlassian's engineering team describes as serving "millions of users." In internal evaluations, programmatic tool calling — which the team calls "code mode" — cut latency of complex Jira-related queries by over 50%, reduced token consumption by 55%, and improved accuracy by 30%, measured against a baseline harness that used standard meta-tools across a representative set of complex queries.
The shift is framed as inevitable: "Every user should have access to an always-on sidekick capable of tackling large tasks, planning work, and managing complex, cross-cutting workflows," the team writes. Where coding agents have traditionally run locally on a user's machine using the local OS as a sandbox, the same design choices now suit generalized personal assistants — pushing Rovo toward persistent, background execution.
Rovo Chat launched as a tool for searching internal Atlassian data, then expanded to retrieve from external systems; it now supports over 50 third-party data sources. The current agent harness is a redesign — splitting the control plane (where the agent lives) from the data plane (where code executes) — replacing the previous synchronous tool-call model.
Five architectural choices anchor the new harness: a split-plane design so sandbox crashes don't kill user conversations, with the control plane able to spin up new sandbox containers transparently; programmatic tool calling inside the sandbox, offloading data iteration and pagination to code execution rather than sequential LLM-driven tool calls; dynamic model selection — direct tool calls for lightweight queries, sandboxed code execution for heavy analysis; a callback bridge that routes every external API call back through the control plane for permission validation, isolating the sandbox from direct network access; and interactive sub-agents instantiated as durable hidden Rovo Chat conversations, capable of outliving their parent's response turn, maintaining independent history, and waking the parent on completion.
The harness supports over 50 third-party data sources, with thousands of API actions exposed through MCP servers and a CI/CD pipeline that dynamically versions and packages tools into a localized Python package that updates the sandbox on warm start in under 100ms. The architecture is designed to scale so that "one scaled service will handle events for millions of users," and the sandbox can be resized mid-execution when compute demand spikes.
The team positions this as the foundation for "a truly autonomous, always-on digital twin" — one that remembers user actions, tracks past conversations, monitors streams, triages notifications, and executes complex workflows autonomously. Self-evolution is the stated flywheel: when the agent encounters a failed query or inefficient path, telemetry is flagged, analyzed, and fed back into the evaluation suite, with engineers reviewing and approving optimizations before they ship.