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

Atlassian’s Long Horizon reasoning engine replaces the previous Hybrid Orchestrator with a unified single-LLM approach supporting extended reasoning across up to 150 iterations, using a flattened tool architecture, progressive tool disclosure, adaptive reasoning depth, and explicit context management through compaction and task decomposition. It demonstrates an 8.5% improvement in offline quality metrics and 23% relative gains in task completion compared to production systems, while emphasizing end-to-end visibility, failure recovery, and efficient token usage through prompt caching.