開源代理應用
OpenMAIC 1.0 Adds an Interruptible Agent Workspace, Moves Course Generation State to Server-Side Persistent Storage
OpenMAIC 1.0 expands its one-shot lesson generator into a course agent that can plan, revise, and resume work, with sessions stored through event streams and PostgreSQL. The new release provides 20 built-in skills and a vendor-neutral tool interface, but public deployments must still add identity isolation and scheduling for long-running media tasks.
The Tsinghua University MAIC team has released OpenMAIC 1.0, introducing a conversation-driven Pro workspace. Users can first have the agent plan a multi-lesson course and then revise slides, quizzes, interactive pages, or project-based learning content item by item. The agent can also read uploaded documents, audio, video, and existing PPTX files, rather than generating a single lesson from one prompt in a one-shot process.
The more significant changes are in the execution layer. The new version moves agent sessions to a background runner and stores conversations as replayable event streams, supporting cancellation, resumption, and additional instructions during execution. Courses, materials, quotas, and sessions can be stored in PostgreSQL. The tool layer covers course planning, page generation, images, video, speech, folders, and web search, and includes 20 skills plus per-owner skill management. Model calls are deliberately decoupled from the content generation pipeline: `@openmaic/generation` only requires callers to supply a function that accepts system and user prompts, without tying it to a specific model provider or storage backend.
This design moves the education agent beyond short-lived demos toward recoverable, long-running workflows, while allowing the course DSL, renderer, and generation pipeline to be reused by other applications. However, the official deployment documentation explicitly warns that the persistent token used for demonstrations does not provide genuine user isolation and must not be used directly for a public service. Browser-side feature flags must also be injected at build time. Recent issue tracking further indicates that video generation still synchronously blocks the agent turn, while its internal 15-minute polling budget is cut short by an outer 10-minute tool timeout. Deployers should move media generation to an asynchronous job queue and add tenant authentication, quotas, source-file permissions, and model cost monitoring before enabling multi-user access.