Long Horizon: How Atlassian Built a Reasoning Engine for Complex AI Tasks

This post introduces the 'Long Horizon' architecture for Atlassian's Rovo AI, replacing the previous hierarchical multi-agent orchestrator. Key technical changes include: a single LLM and context for complex tasks, adaptive reasoning depth based on query complexity, a flattened tool architecture where all product capabilities are exposed as uniformly named actions (e.g., jira__search_issues), and progressive disclosure of tool schemas via meta-tools ({product}__get_tool_schema, {product}__invoke_tool) to manage cost and accuracy. It also details context window management through a Context Compaction Service and task decomposition via child instances for wide tasks. The previous architecture's limitations in iteration depth, context sharing, and model adaptation are addressed by this new design.