代理框架
JarvisHub Turns the Canvas Into Shared State for Multimodal Agents, but Its First Release Still Lacks Quantitative Evaluation
The open-source JarvisHub uses typed nodes and edges to preserve assets, versions, dependencies, and agent trajectories, allowing long-horizon creative work to live outside linear conversations. The system already provides an executable three-layer architecture and local deployment tooling, but the paper currently presents only qualitative case studies and has yet to demonstrate improvements in reliability or output quality.
JarvisHub was open-sourced on July 28, positioning an editable canvas simultaneously as a human–AI workspace, external memory for agents, a protocol-constrained action space, and persistent state for multimodal projects. It targets long-horizon creative work involving images, videos, websites, and presentations. Such work encompasses not only final outputs but also reference materials, candidate versions, failed attempts, localized edits, feedback, and cross-asset dependencies—all of which are difficult to represent precisely in a linear chat history.
The framework is divided into three layers. Canvas State uses PostgreSQL, a shared schema, and a layout package to store editable nodes, spatial arrangements, versions, and status information. Protocol Bridge exposes tool capabilities and execution permissions, validating canvas edits before committing state transitions. Agent Runtime reads the current canvas, invokes skills, memory, tools, or sub-agents, and then returns observations and artifacts that can be committed. Text, character settings, storyboards, code, previews, and revision feedback can all be represented as typed nodes, while edges can express references, generation dependencies, and version lineage.
The open-source release includes a web interface, Hono API, agent bridge, and trajectory viewer, with support for image, video, audio, browser, document, presentation, and MCP tools. The default development environment requires Node.js 24.15, pnpm, Python 3.10 or later, and PostgreSQL; Docker Compose is also provided. The Apache 2.0 license lowers the barrier to research and derivative development.
However, the authors explicitly acknowledge that the current release presents only three categories of qualitative case studies—narrative media, interactive websites, and presentations—with no completed benchmarks or controlled experiments. Final output quality also remains dependent on the external models used. In addition, the default shared test bucket may allow other users to read, overwrite, or delete objects, so private data should not be uploaded. Key areas to watch next include state-conflict handling, failure-retry semantics, the cost of long trajectories, and quantifiable task-completion rates compared with chat-based and node-based workflows.