AI Agents Vibe Coding Course / GoogleとKaggleの5日講座を試す

1 English
2 日本語
3 References / 参考リンク

English

Google and Kaggle are bringing back their free five day AI Agents Intensive course from June 15 to June 19, 2026. The timing is good. A lot of people want to build agents, but many still get stuck at the same point: they can chat with a model, yet they are not sure how to turn that into a repeatable workflow that actually saves time.

The official announcement is short, but it gives a useful outline. The course is online, free to registrants, updated from the earlier version, and ends with a capstone project. Google also says the first run reached more than 1.5 million learners. That scale matters because it suggests the course is not being framed as a niche research seminar. It is positioned as practical training for people who want to build real agent systems.

Why this course is worth watching

The most important line in the announcement is not the marketing language. It is the idea that natural language becomes a primary programming interface. That is the part many people care about right now. You do not need to throw away your editor or your existing workflow. You need a way to describe work clearly enough that an agent can plan, call tools and hand back something usable.

Google also says the course covers building production ready systems and connecting tools and APIs. That is the difference between a toy demo and something you might actually keep using. A decent agent is not just a chat window with a fancy label. It needs boundaries, inputs, tool access and a format for output that does not create cleanup work for you later.

My view is simple: if you want to understand agents, start with one ugly repetitive task from your own week. Do not begin with a grand personal assistant. Begin with a workflow that already annoys you.

A realistic first project

If you take the course, a good first project is an inbox triage agent for work. The job is narrow enough to be testable, but useful enough to feel real.

Input: pasted emails or support requests
Decision: label each item by urgency, owner and next action
Output: a short table or checklist you can review in under two minutes

This is a better learning target than asking an agent to run your whole job. You can measure whether it is helping. You can also spot failure quickly, which is what you want while learning.

A prompt pattern you can borrow

You do not need a huge prompt library on day one. A compact instruction with a fixed output shape is enough to learn the basics.

You are an operations triage agent.
Read the message set below.
For each item, return:
1. urgency: low, medium, or high
2. owner: sales, support, finance, or me
3. next_action: one sentence
4. draft_reply: optional, only if the request is routine
Keep the output in a Markdown table.
If information is missing, say what is missing instead of guessing.

The point is not that this exact prompt is special. The point is that it gives the agent a role, a schema and a refusal rule. That combination usually matters more than making the wording sound clever.

How to use the five days well

If you register, I would treat the course like a working sprint instead of passive content. Day one should end with one tiny workflow. By the middle of the week, connect one external tool or API. By the final day, make the capstone reflect a task you already do, not a made up showcase.

A practical rhythm could look like this:

Day 1: pick one recurring task and define the output format
Day 2: test prompts with five real examples
Day 3: add a tool call, file step or API step
Day 4: tighten failure handling and edge cases
Day 5: package it into a repeatable routine you can reuse next week

That is the kind of pace that teaches you something durable. Watching agent demos is easy. Keeping one small agent alive for a week is where the real understanding starts.

Cost and who this fits

The course itself is free according to Google’s announcement. The real cost is your attention. If you try to absorb everything at once, you will probably remember almost nothing. If you enter with one concrete problem, the course becomes much more valuable.

This topic fits people who want to learn agent workflows for coding, operations, research or internal coordination. It is also a good fit if you have already tried ChatGPT, Gemini or Claude for one off tasks and now want something more structured than ad hoc chatting.

If I were blocking time for this, I would reserve one short session each day and keep the scope brutally small. That is usually how a useful agent starts: not as magic, but as one annoying task that finally stops eating your morning.

日本語

GoogleとKaggleが、無料の5日間講座「AI Agents Intensive」を2026年6月15日から19日に再開催すると案内しています。こういう講座って聞くと大げさに見えがちですが、今回はかなり実務寄りです。AIエージェントを触ってみたいのに、「結局どこから作ればいいのか」が曖昧な人にはちょうどいい入口です。

公式発表で押さえておきたいのは、前回実施時に150万人超が参加したこと、今回は内容更新と新しい登壇者が入り、最後にキャップストーン課題まで用意されていることです。単なる概念紹介ではなく、実際に作る前提の流れになっています。

この講座の見どころ

いちばん大事なのは、「自然言語を主なプログラミングインターフェースとして扱う」という考え方です。つまり、コードを1行ずつ書く前に、まず人間の言葉で作業の流れを定義し、その指示をもとにエージェントへ仕事をさせる発想ですね。

しかもGoogleは、基礎から本番投入レベルまで扱うと書いています。ここが重要です。エージェントは会話できれば終わりではありません。入力をどう受けるか、どのツールを呼ぶか、出力をどの形式で返すかまで決めて、ようやく業務に乗せられます。

個人的には、こういう学習でいきなり「万能秘書AI」を目指すのはおすすめしません。最初はもっと地味でいいです。毎週くり返す面倒な作業を1つだけ切り出した方が、途中で迷いません。

最初の題材は inbox 整理ぐらいがちょうどいい

講座で試す題材として現実的なのは、受信メールや問い合わせの振り分けです。範囲が狭いので評価しやすく、しかも使い道がはっきりしています。

入力: メール本文や問い合わせ文を貼る
判断: 緊急度、担当、次に取る行動を決める
出力: 2分以内に見直せる表やチェックリストを返す

これなら、役立っているかどうかを自分で判断できます。うまくいかない時も、どこが悪いのか見つけやすいです。学習初期はこの「壊れ方が分かる」ことがかなり大事なんですよね。

講座中にそのまま使える指示の型

初日から複雑なプロンプト集を作る必要はありません。役割、出力形式、推測しすぎないルール。この3つだけでもかなり形になります。

You are an operations triage agent.
Read the message set below.
For each item, return:
1. urgency: low, medium, or high
2. owner: sales, support, finance, or me
3. next_action: one sentence
4. draft_reply: optional, only if the request is routine
Keep the output in a Markdown table.
If information is missing, say what is missing instead of guessing.

この例で大事なのは、表現の巧みさではありません。役割を固定し、返却フォーマットを決め、分からない時の挙動まで指定している点です。エージェント作りでは、気の利いた言い回しより、この骨組みの方が効きます。