Andrew Ng has published a practical guide to building an AI agent harness, alongside the open-source OpenWorker Harness SDK. The useful takeaway is that a coding agent can assemble a working demo, but the tools, permissions and approval steps still need careful boundaries.
Andrew Ng Watch analysis
What happened
Ng’s guide explains an agent harness as the code that gives a language model context, tools and guardrails. It describes using a coding agent to build a file-organising demo with OpenWorker’s SDK.
The example starts with npx openworkerai/openworker-harness-builder –agent claude –project. Ng says readers can substitute codex for claude when using Codex. The demo gives its agent tools to list and read files, create or delete subdirectories, and move files.
Our top picks
- An open-source harness SDK
OpenWorker’s SDK is presented as a way to build agents with different model providers and deployment options. - A coding-agent-built demo
The guide walks through creating a local agent to organise sample files, rather than stopping at a conceptual diagram. - Tools with a fenced work area
The example restricts file operations to a dedicated demo directory, limiting where the agent can act. - Approval before changes
The demo requires user approval before each subdirectory creation and file move. - Visible agent activity
It displays tool activity and a short status of the agent loop, so readers can follow what it is doing.
Why it matters
The harness is the layer that turns a model call into an application that can use tools and take steps. That makes it important practical infrastructure: a capable model is only part of an agent, and the surrounding code determines what it can access and how its actions are controlled.
Ng’s example is deliberately small, but it gets the essentials on the table: restrict file access, keep credentials away from the model, provide a way to approve consequential actions, and consider sandboxing. The demo is a starting point, not evidence that every agent built this way is secure.
Our read
This is a useful, hands-on guide to the plumbing behind agents, with sensible controls built into the example rather than bolted on as an afterthought. If you are experimenting, start in a disposable project directory, keep permissions narrow and make the agent show its work. “It seemed fine” is not a security review, however charming the demo.
What to watch
- Whether developers find the SDK easier to adapt than provider-specific harness tools.
- How the project handles credentials, sandboxing and other controls in more complex agents.
- Whether the SDK gains users beyond the demo and its initial examples.
Discussion spark: For an agent that can change files, should human approval be required for every action, or is a tightly restricted sandbox enough?
Sources and evidence
- Source update (13 November 2026, 15:30 UTC)
Independent WittyWires tracker for public updates about Andrew Ng. Not affiliated with or endorsed by Andrew Ng; this is not an official account.