SYSTEM Cited by 1 source
Atlassian Long Horizon¶
Atlassian's Long Horizon is a reasoning-native agent orchestrator that powers Rovo Chat. It replaced the Hybrid Orchestrator (a hierarchical multi-agent system with per-product sub-agents) with a fundamentally different design: one LLM, one context, one iterative loop running up to 150 iterations per user turn.
Architecture¶
- Single reasoning loop. One LLM holds the full conversation state and all available tools. No intermediate agents paraphrase results. The model that decides to make a tool call is the same model that reads its response and decides what to do next.
- Flattened tool surface. Every product operation (Jira,
Confluence, Bitbucket, JSM, Compass, plus third-party connectors)
is a typed, namespaced action (
jira__search_issues,google_calendar__list_events) callable directly. - Progressive disclosure. Each product namespace exposes two
meta-tools (
get_tool_schema+invoke_tool) with a one-line summary of all tools in the namespace description. Frequently-used tools stay permanently flat. Schema cost: fetched once per tool per task. - SKILL.md per namespace. Hand-authored domain guides encoding which tool to reach for, multi-step recipes, and gotchas — replacing the implicit expertise that previously lived in sub-agent prompts.
- Context compaction service. Runs before each model call; trims or summarises older tool outputs while preserving recent results at full resolution. Pruned content is offloaded for on-demand retrieval.
- Child instances. For wide tasks, spawns parallel Long Horizon instances — each with its own clean context — to own independent research strands. Parent synthesises finished results.
- Prompt layer ordering. Assembled from most-stable to
most-volatile for maximal prefix-cache reuse. Explicit
cache_controlmarkers for Anthropic; implicit caching for OpenAI/Gemini. - Adaptive reasoning. Calibrates reasoning depth per query (minimal for lookups, deep for multi-step research).
Production numbers¶
| Metric | Value |
|---|---|
| Iteration budget | Up to 150 |
| Timeout | 20 minutes |
| LLM calls per tool use | 1 (vs 2 in Hybrid Orchestrator) |
| Offline accuracy | 77% (+8.5% vs previous) |
| Task completion (Confluence) | +23% relative |
| Chat success rate (online A/B) | +0.83% |
| Perceived latency | −37% (streaming progress) |
Predecessor: Hybrid Orchestrator¶
The Hybrid Orchestrator was a hierarchical multi-agent system (coordinator + per-product specialist sub-agents). Limitations that motivated Long Horizon: - Information loss. Orchestrator only saw summaries from sub-agents; never raw tool outputs, intermediate reasoning, or errors. - Limited iteration depth. Optimised for 2–5 step tasks; multi-step research hit the ceiling. - Models outgrew the box. Frontier models handle longer contexts and larger tool surfaces reliably; fragmenting context across sub-agents prevented them from using these capabilities. - N-way model migration. Every model upgrade required re-tuning N sub-agents. Long Horizon: one model migration, not N.
Seen in¶
- sources/2026-06-18-atlassian-long-horizon-reasoning-engine — canonical architectural description.
Related¶
- systems/rovo-chat — the user-facing product Long Horizon powers
- systems/rovo-dev — sibling Atlassian AI agent (development)
- concepts/context-window-as-token-budget
- concepts/context-compaction
- concepts/progressive-tool-disclosure
- patterns/single-loop-agent-orchestration
- patterns/flattened-tool-architecture
- patterns/context-compaction-service