HKUDS Releases Multi-Agent Runtime

The HKU Data Intelligence Lab (HKUDS) has released OpenOPC. This runtime automatically generates an org chart containing necessary roles and reporting structures upon goal input, and deploys AI employees to execute tasks, addressing the context limitations of single agents and the inefficiencies of human coordination.

Core Operating Principles

OpenOPC operates on three axes: Self-Built, Self-Run, and Self-Grown.

• Self-Built: Derives necessary roles and dependencies from goals as a graph to create an org chart, deciding whether to reuse existing employees or hire new ones.
• Self-Run: Manages collaboration through a Work Item state machine with five modes: execution, delegation, review, integration, and rework. Independent tasks are processed in parallel, while dependent tasks are handled in a waiting state.
• Self-Grown: Distills execution records into role-specific lessons stored in personal experience profiles, promoting recurring lessons to shared playbooks.

Technical Features and Use Cases

• Compatibility: Runs on Python 3.10 or higher, allowing external coding agent CLIs such as Codex, Claude Code, and Cursor to be assigned per role. It is not tied to specific model providers via LiteLLM and OpenRouter.
• UI and Approval Policies: Provides terminal commands and a browser-based Office UI. Security measures include classifying risk levels for shell command execution, always escalating destructive commands to humans, and subjecting medium-risk commands to LLM review.
• Suitable For: Teams that need to repeatedly perform long tasks where multiple stages wait on each other, and who want to visually track the process while leaving room for intervention. Note that current CLI features may be limited compared to the UI.