Google's Two-Hour Course Lays Out the Full Agent Roadmap—but the Comments Pour Cold Water First
The course goes from Prompt to Self-Improving Graphs in a two-hour video. The real issue is in the comments: without stopping conditions and evals, self-improvement turns into self-drift.
Google released a two-hour course on X with a very straightforward title: From Prompts to Agents, then Harness, Loops, and Self-Improving Graphs. The original post was reposted by RoundtableSpace and credits the author 0xMarioNawfal. The video is two hours long—not a few-minute concept piece.
Let’s lay out the terms first. A Prompt is the task instruction given to the model; an Agent is an executor that can call tools and break down steps; a Harness is the runtime framework wrapped around the model; Loops make execution run repeatedly; and Self-Improving Graphs let the process change its structure based on results. For people with a development background, these terms aren’t new. For ordinary users, the core idea comes down to one sentence: the system has to run, it has to be inspectable, and it has to be stoppable.
The course arranges the roadmap as an upward line. The comments section first asks a more down-to-earth question: who grades the work? A user named Mohammad Vazin says in a reply that going from harness to loops is where most teams get stuck. The quality of a loop depends on stopping conditions and evals. Without a cheap check to determine whether this output is worse than the last one, self-improving quickly becomes self-drifting. That judgment is closer to real engineering than the course title.
Another reply comes from Amy. She says self-improving graphs sound fancy, but she’d rather have an agent that can complete a task without someone watching it. It’s not a pleasant thing to hear, but it’s accurate. Getting something to run once in a demo and getting it to run reliably one hundred times in production are separated by evals, retries, logs, and rollbacks.
The Floor Trading Club’s reply is shorter: Prompts are just the starting line. When a course goes from Prompt to graphs, it’s easy to think the focus is the final graph. In real deployment, first check whether the loop has stopping conditions, whether evals are cheap enough, and whether failures can be rolled back. Without these, self-improving just runs errors for longer.
Original post: https://x.com/RoundtableSpace/status/2116609902798123239
发布时间: 2026-11-01 22:43