Opening the Door to Agent Autonomy: The Architecture Behind Rovo’s Agent Harness
Atlassian’s Rovo Chat expands beyond data search to support over 50 third-party sources and introduces an agent harness architecture to enable end-to-end workflow automation. The new system separates data and control planes for scalability and reliability, using isolated sandboxes and asynchronous multi-agent coordination to power always-on digital assistants.
Rovo Chat initially focused on retrieving internal Atlassian data but now integrates over 50 external sources, positioning it to handle end-to-end user workflows. The platform’s evolution reflects a broader industry trend where coding agents and chat applications converge, enabling more generalized personal assistants. To achieve this, Atlassian redesigned its execution model with a distributed architecture that separates the data plane from the control plane, allowing independent scaling and improved resilience. This split ensures reliability even if one component fails, abstracting sandbox failures from the user experience.
The agent harness uses isolated compute sandboxes and asynchronous multi-agent coordination to maintain continuous operation. Sandboxes scale independently from the control plane, enabling lightweight interactions for simple queries or robust environments for complex tasks. By separating the sandbox from the user interface, the system ensures seamless recovery from failures without disrupting the user. This design allows millions of users to share a single scaled service while enabling agents to respond instantly, increasing operational density.
The architecture supports programmatic tool calling, offloading data iteration and pagination to code execution for efficiency. This approach, similar to implementations by Anthropic and Cloudflare, reduces latency and token consumption while improving accuracy. Tools are dynamically versioned and packaged via CI/CD pipelines, updating sandboxes on warm starts to maintain sub-100ms startup times. The system validates user permissions and authenticates requests before external API calls, ensuring secure tool execution within isolated environments.
A multi-agent system enables parent agents to delegate work to persistent sub-agents that operate independently in the background. These sub-agents maintain their own conversation history and can notify parent agents upon task completion, avoiding the need to reinitialize context for follow-up actions. This design supports parallel workflows without overburdening the primary agent, enhancing scalability and reducing latency for complex queries.