Andrew Ng’s OpenWorker: An Out‑of‑the‑Box AI Agent Built for Getting Real Work Done

Andrew Ng (吴恩达) recently announced an open‑source AI agent product called OpenWorker . Its tagline is “no chit‑chat, just get work done”: users state the desired result, the agent decomposes the task, calls the appropriate tools, and presents the finished output such as a document, a Slack message, or an updated calendar entry.

The design makes several deliberate choices:

Model‑agnostic architecture : OpenWorker can work with any LLM, including GPT‑5.6 Sol, Claude Fable, Gemini 3.6, as well as open‑source models like Kimi, GLM, DeepSeek, or locally‑run Ollama instances. The model can even be switched mid‑conversation.

Data privacy : Only the selected LLM and integration tools see the data; all other processing—including the agent loop, dialogue state, tokens, and API keys—remains on the user’s machine.

Human‑in‑the‑loop approval : Before performing any action that modifies external resources (sending email, editing a calendar, executing commands), the agent pauses and asks for confirmation. This addresses concerns such as “the agent might delete all my meetings when it gets bored.”

Four ready‑to‑use scenarios are highlighted:

Sales – research customers and draft emails.

Administrative assistant – manage calendars and clean inboxes.

Marketing – track spend and generate weekly reports.

Operations on‑call – handle alerts and write incident records.

OpenWorker integrates with more than 25 tools, including Slack, GitHub, Jira, Notion, Linear, HubSpot, Outlook, Gmail, and others via the MCP protocol, and can be extended further.

The system is built on aisuite , an open‑source unified LLM interface library previously created by Andrew Ng’s team. Developers who want to build their own agent framework can start from aisuite; OpenWorker serves as a complete reference implementation.

Andrew Ng’s broader view, expressed in recent talks, is that the next breakthrough in AI is not larger models but agents that can actually accomplish tasks. OpenWorker is presented as the code embodiment of that insight.

Repository: https://github.com/andrewyng/openworker