Breaking

Andrew Ng’s Agent Harness Push Reframes Prompting as AI Systems Engineering

September 27, 2026 at 12:45 EDT

AI Agents
Software Dev & Coding
Open Source

Andrew Ng is pushing developers to move beyond one-off prompts toward agent harnesses: the software layer around large language models that supplies tools, memory, guardrails and feedback loops. The framing turns a viral debate about the future of prompting into a practical systems question: how to make agents reliable enough to do work, not just answer chats.

2026 · Agent engineering

From prompts to harnesses

Andrew Ng’s latest agent framing puts the focus on the software around the model: tools, memory, approvals, logs and feedback loops that turn chat into completed work.

Course

7h45m

Agentic AI instruction

Lessons

Video lessons

Practice

Code examples

Reach

25+

OpenWorker integrations

Course structure, drawn to scale

Video lessons

Assignments

Code examples

The training emphasis shifts from wording a request to building loops that can plan, use tools, evaluate results and recover.

Why builders like it

Harnesses preserve context, call tools, log actions and can turn a request into a finished file or workflow.

What still limits it

Complex agents add overhead, need evals and observability, and should keep humans in control for risky actions.

The new build loop

Prompt→Tools→Harness→Verified work

Prompts remain inside the system, but reliability moves to the surrounding runtime.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.