SYSTEM Cited by 3 sources

Rovo Chat

Atlassian's Rovo Chat is the user-facing AI assistant product that lets Atlassian customers ask questions, run research, and take actions across Jira, Confluence, Bitbucket, Jira Service Management, Compass, and third-party connectors (Google Calendar, Google Drive, Slack, GitHub, Microsoft Teams).

Architecture evolution

1. Hybrid Orchestrator (2025) — hierarchical multi-agent: coordinator LLM decomposes query → routes subtasks to per-product sub-agents (JiraAgent, ConfluenceAgent, BitbucketAgent, etc.) → combines summaries. Optimised for 2–5 step tasks. Limited by lossy hand-offs, shallow iteration budget, conservative timeouts.

2. Long Horizon (2026) — single-LLM, single-context, iterative reasoning loop. Up to 150 iterations, 20-minute timeout. Flattened tool surface + progressive disclosure + context compaction + child instances for parallel research. See systems/atlassian-long-horizon for full architecture.

Capabilities

• Cross-product research (Jira + Confluence + Slack + Bitbucket + ...)
• Multi-step planning and execution
• 14+ pre-authored skills for common task types (sprint planning, bug triage, performance review, competitive analysis)
• Adaptive reasoning effort (simple lookups vs deep research)
• Real-time streaming progress updates during reasoning
• Task decomposition into parallel child instances

Execution runtime (2026-08-27)

Beneath Long Horizon's reasoning loop sits the Rovo Agent Harness, the distributed runtime that actually executes the agent's decisions. It moves Rovo Chat from a synchronous data-retrieval chatbot toward an "always-on sidekick" via a split-plane architecture (durable control plane + disposable on-demand sandbox), programmatic tool calling / code mode in the sandbox (code-generation-over-tool-calls — measured >50% latency, 55% token, +30% accuracy vs a meta-tool baseline on complex Jira queries), dynamic model selection (skip the sandbox for lightweight queries), a callback bridge that routes sandbox tool calls back through the control plane for auth, interactive async sub-agents (a sub-agent is a hidden Rovo Chat conversation with a durable handle), discovery-probe + checkpoint error recovery, and a human-approved self-evolution loop. (Source: sources/2026-08-27-atlassian-opening-the-door-to-agent-autonomy-the-architecture-behind-rovos-agent-harness)

Seen in

• sources/2026-06-18-atlassian-long-horizon-reasoning-engine — canonical Long Horizon architecture post.
• sources/2026-08-27-atlassian-opening-the-door-to-agent-autonomy-the-architecture-behind-rovos-agent-harness — the agent harness / runtime beneath Rovo Chat.

Related

• systems/atlassian-long-horizon — the reasoning engine powering Rovo Chat
• systems/rovo-agent-harness — the distributed execution runtime beneath Long Horizon
• systems/rovo-dev — sibling AI agent for development workflows
• systems/atlassian-ml-studio — platform serving Rovo models

Event-guide deployment (2026-08-07)

The Atlassian Unleash PWA used Rovo + Confluence as the event-guide surface: attendees asked the guide questions while the PWA retained Jira for operational event state. The post does not disclose the grounding, authorization, or tool-execution mechanics, so this is evidence of product composition rather than a new Rovo Chat architectural claim. (Source: sources/2026-08-07-atlassian-building-a-real-time-pwa-on-atlassians-own-stack)