Creating with Rovo: How We Built a Collaborative AI Canvas

This post details the technical implementation of a collaborative AI canvas for content creation within Atlassian's Rovo AI solution. It covers the design principles of creation living inside chat and being a platform, not a feature, with real-time streaming. Key technical contributions include: 1. A new Confluence creation and editing skill agent for Rovo. 2. LLM output formats for different content types: ADF (nested JSON) with a repair library for pages, SVG for whiteboards (converted to native elements), and CSVs (schema, views, data) wrapped in XML for databases. 3. Agentic loops with reflection for page editing using editor-style commands. 4. Real-time SVG parsing and constraint solving for whiteboard assembly. 5. A streaming API contract for LLM actions and the Rovo Bridge API for bidirectional communication between chat and content objects, abstracting transport layers (BrowserTransport, DesktopTransport) and using JSONRPC over the chosen transport. 6. Offline evaluations using screenshot-based assessment and online evaluations tracking success rates, alongside automated health checks for agentic flows.