Wink Pings

Google released a free 2-hour Harness engineering course, and the most expensive part of paid courses has been covered

The course covers agent building, eval engineering, loop engineering, and self-improving graph systems from scratch. The poster says it is worth more than a $500 course, and someone in the replies warns: eval is the part most people skip.

Google has just released a free 2-hour Harness engineering course, covering everything from the first step to the final stage. The poster rob. listed the timestamps on X and said he had seen some $500 courses that were worth less than this video.

Here are the timestamps first:

  • 0:29 Harness engineering from scratch
  • 11:08 your first agent, properly built
  • 28:08 Eval engineering with a “Chief of Staff” agent
  • 1:00:13 Loop engineering
  • 1:23:19 a graph system that improves itself and never stops

To put these sections in plain terms: understand Harness engineering from scratch; build your first agent properly; use a “Chief of Staff”-style agent for eval; cover loop engineering; and finally land on a graph system that can self-improve and keep running. In two hours it connects the four things—agent, eval, loop, and graph—and the information density is not low.

The term Harness engineering has not yet been worn out the way prompt engineering has. You can understand it as the engineering framework outside the model: how to organize tools, tasks, evaluations, and feedback so that an agent works reliably in real workflows. Model capability is one layer; whether you can connect that capability into a repeatable system is another. Many courses stop at getting a demo to run, while this course pushes directly from agent building to eval and loop; the order feels more like an engineering course than a marketing course. The most expensive part of paid courses is often not how to build, but how to judge whether you built it right; that part appears at 28:08.

A few judgments in the replies are worth noting. Kevin Moreira said the idea of using a “Chief of Staff” agent for eval is very good, and he had not seen anyone frame it that way before. Fenix was more direct: eval is the part everyone skips, and skipping it leads to trouble. Marc Kessner thinks loop engineering plus evals is the combination that creates separation. Stefan Friedman calls Harness engineering the most in-demand skill of 2027; that prediction can be treated with skepticism, but the direction is not far off.

Some readers also asked practical questions: do you need to know Python, and can you follow along without much background? The material does not provide an answer. Judging by the course structure, the first two sections should be understandable conceptually, but by the eval, loop, and graph-system parts, it is best to have some experience with code and agents; otherwise it is easy to just watch without really following.

The poster also mentioned that the course is free. Someone in the replies said, “Content at this level for free—times have changed,” and someone else said they would get their team together to watch it on Friday. For AI practitioners with a development background, the highest value is the eval section after 28 minutes and the loop section after one hour; these two parts determine whether an agent can go from a demo to a system. For non-technical readers, the first 30 minutes can serve as a map of agent engineering—at least you will know that an agent is not just a model plus prompts.

Original post: https://x.com/robiartec/status/2123055373737148722

If you only have two hours this weekend, watch the eval and loop sections first, then go back and fill in the from-scratch part. Don't rush to buy a course; first watch the free one to align on terminology and workflow.

发布时间: 2026-11-19 07:31