JarvisHub: A Canvas-Native Framework for Long-Horizon Multimodal Creative Agents
新智元 · wechat · 2026-07-30
Several top universities jointly introduced JarvisHub, a canvas-native harness for long-horizon multimodal creative agents. While existing tools often discard intermediate assets, JarvisHub explicitly saves the entire project state as an editable multimodal asset graph, enabling agents to continuously advance complex tasks.
- Canvas as State: Prompts, reference images, and version dependencies are saved as stable nodes. Agents can precisely target materials for local modifications without starting over.
- ProtocolBridge: Acts as a security layer between the agent and the canvas, ensuring every modification is authorized and validated to prevent accidental deletion or overwriting.
- Unified Runtime: Integrates tool orchestration, memory, and sub-agents. It supports local failure recovery and feedback trajectory recording, providing structured data for future agent benchmarks and trajectory training.
More from coding & agent
- agensis Launches AI Agent Collaboration Workspace with Self-Hosting and Unlimited Agents — jasonkneen · 2026-07-30
- Stop Copy-Pasting: Connect AI to Team Knowledge via MCP — femke_plantinga · 2026-07-30
- Developer releases smolagents adapter to use ChatGPT Codex OAuth instead of API key — woldhack · 2026-07-30
- AI coding agents make building software easier than understanding it? Developers discuss — bluetech333 · 2026-07-30
- What Did the AI Do While You Slept? Dev Shares Overnight Agent Logs — Dimillian · 2026-07-30
- Building Local AI Voice Agents: vLLM + Kokoro + Livekit — ashishllm · 2026-07-30