シャドーイング練習: 4 Types of AI Agent Loops: From Human-Led to Autonomous - 動画で英語スピーキングを学ぶ

練習モードを選ぶ

各モードは全画面の練習ルームで開きます。

95 文
文が短すぎたり長すぎる場合は、Editをタップして調整してください。
1
AI agents work autonomously and they do so using loops
2
but the level of autonomy and the form of the loop well they can differ.
3
So let's take a look at four types of agentic loop and why you might opt for each one.
4
So at a high level all agentic loops they share the same basic patterns.
5
So we have an agent and the agent takes in a goal.
6
That is the goal it's going to meet.
7
Then it performs an action to help it achieve that goal.
8
That might be something like editing a file or maybe running a command.
9
The next stage in the loop is really that it observes what comes back from the action,
10
so what happened, maybe like reading some of the output.
11
And then the next stage in the loop is that it adjusts based upon what it has observed,
12
and it kind of goes round and round again until the job is done
13
so two things about that who sets the goal and who decides
14
when the loop's done
15
when the goal is achieved well it's often a human a human in the loop
16
so to speak not always but often
17
so let me introduce the first of the agentic loops i want to cover the interactive loop
18
so the interactive loop is this loop cycle where human involvement is going to be added in in two places.
19
So first of all, we have a human here at the start,
20
and the human's job at the start is to set the goal.
21
What do we want the loop to achieve?
22
So the human sets the goal.
23
Now, in practice what this means is the human types in a prompt specifying what they want
24
and the human judges at the end as well
25
so we've got another human step here where the human judges what has been produced
26
and they look at this to see is the job done what came back is the goal achieved
27
so if the goal was to upgrade library x then
28
if the upgrade happened then well the human would judge the loop as being
29
done if it wasn't or
30
if there's more to do the human is going to type the next prompt
31
and round we go again so the agent runs inside of the loop
32
and a person controls when the loop starts and when the loop ends.
33
So if the task's a bit fuzzy or there's not really a measurable finish line,
34
an interactive loop with a person controlling start and stop conditions is probably a good idea.
35
But the interactive loop is hard to scale.
36
When the job is basically doing the same thing over and over again, well these poor humans will have a lot of judgments to make.
37
So what if we handed
38
that judging job off to the ai well
39
that's the second type of loop the delegated loop the human still sets the goal
40
but the goal comes with a little extra here the goal comes with a stopping condition simply
41
when should the loop actually be done
42
so a goal isn't going to be upgrade library x
43
it's going to be something a bit more specific like upgrade the out of date libraries
44
and stop when the build turns green now at the end of the loop we now have a check
45
that we go to after adjust and the check is going to judge this stopping condition and it's an ai check
46
So every time the agent thinks it's finished after this adjust stage, the check is run.
47
And if it fails, then the agent will go around the loop again.
48
And if it passes, the loop now exits, comes out of the check because we think the goal has been completed.
49
And now it goes to another stage.
50
Basically the result comes back and it is provided for a human to review
51
the result of that job so the delegated loop here delegates the judging away from a human
52
until it's time to review the finished work so like go off
53
and do that task i sent you mr agent
54
and let me know when you're done then i'll see how well you did
55
but a delegated loop still has a human setting the goal
56
and starting the loop so what if we handed that off as well well that's the third loop, the scheduled loop.
57
Most everything from the delegated loop stays the same in
58
that there's a check that's going to judge the stopping condition and the result comes back to a human to review.
59
But what does change here is how this loop gets started because the thing that starts it now is not a person, it's a clock basically.
60
It's a human writing a goal once, something like find any out -of -date libraries and upgrade them,
61
and then a schedule is going to run it, say every night.
62
So a scheduled loop suits work that goes stale over time, like libraries going out of date,
63
but it's also really good for work that needs doing recurrently.
64
The human's job is to read what comes back, and when this loop stops being useful,
65
to then, well, cancel the loop when we don't need to keep scheduling it anymore or to adjust the schedule as needed.
66
So now the clock starts in place of a person having to kick the loop off, but there's still a human reviewing every result.
67
Yeah, I wonder if we handed that one off as well.
68
Well then we have the fourth type of loop, the event driven loop.
69
And a couple of things change here.
70
So first the clock is replaced with something else it's replaced with an event
71
so the loop only runs
72
when something happens an event like a pull request has been opened
73
and each event is going to kick off its own run
74
of the loop with the goal tailored to whatever the event
75
was about now there's still an AI powered check to decide
76
when that run is done but now the human review at the end has gone away
77
and it's replaced with an AI option basically another agent a reviewer agent that now does the checking.
78
Now this is a second agent with its own fresh context maybe powered by a different AI model
79
and it looks over the finished work before it moves on.
80
Now if it approves the work is done
81
so we can say the loop is complete and
82
if there's still more to do we're in the not done stage
83
and we would go back around the loop again the loop gets restarted
84
and if the reviewer agent isn't sure what to do
85
if it doesn't know if it's done or not done well then that's
86
when we go to a human review stage so at this point the human will only see certain cases to review.
87
They will effectively only see exceptions at this point.
88
So this is the most autonomous of the four.
89
It's still got a review step but it's usually an agent doing most of the reviews that are not the exception cases.
90
So that's four agentic loops.
91
In an interactive loop a human starts the loop and judges every turn.
92
A delegated loop keeps the human at the start
93
but a check does the judging a scheduled loop has a clock to start it
94
and an avenge driven loop is started by whatever just happened with an agent reviewing the result
95
which agentic loops are you running today let me know in the comments

アプリをダウンロード

流暢に話すために必要なすべて

AI発音採点文ごとに採点
IPA練習すべての音をマスター
単語学習自分の単語帳を作る
単語ゲーム遊びながら覚える

同じチャンネル

FOR YOU

おすすめ動画

このレッスンの語彙とスピーキングのポイント

このスピーキングレッスンは、動画「4 Types of AI Agent Loops: From Human-Led to Autonomous」を教材にしています。 繰り返し出てくる語は次のとおりです:loop, goal, agent, review, judge。 この動画には、シャドーイング用の文が95文、単語が1272語あります。 音声の長さは8:56です。 話す速さは1分あたり約142語で安定しており、シャドーイングしやすいペースです。 単語の84%は英語の頻出3,000語に含まれます。残りは練習の前に確認しておきましょう。

この動画の重要語彙

動画に出てくる覚えておきたい単語15語を、発音と意味つきで紹介します。

単語発音意味
loop 名詞/luːp/ループ
delegate 名詞/ˈdɛlɪɡət/代表, 使節
clock 名詞/klɑk/走行距離計
adjust 動詞/əˈd͡ʒʌst/調整する
exception 名詞/əkˈsɛpʃən/除外
observe 動詞/əbˈzɜːv/守る, 従う
output 名詞/ˈaʊtpʊt/出力
introduce 動詞/ˌɪntɹəˈdus/紹介する, 引き合わせる
exit 名詞/ˈɛɡzət/出口
edit 動詞/ˈɛdɪt/編集する
judgment 名詞/ˈd͡ʒʌd͡ʒ.mənt/判決, 裁き
cancel 動詞/ˈkæn.sl̩/取り消す, 取り止める
approve 動詞/əˈpɹuːv/承認する, 賛成する
differ 動詞/ˈdɪfɚ/異なる, 違う
autonomy 名詞/ɔˈtɑ.nə.mi/自治

動画に出てくる句動詞

単語意味
come back 動詞戻る, 帰る
go away 動詞去る, 出る
go back 動詞戻る, 帰る

この動画の文法

話し手がよく使っている文型を、動画の実際の表現とともに紹介します。

文型動画での表現
条件文 if + 節, will/would + 動詞 — 条件とその結果if it fails, then the agent will go
受動態 be + 過去分詞 — 誰がするかより、何が起きるかに焦点を当てるis achieved · be added · has been produced
現在完了形 have/has + 過去分詞 — 過去の出来事が今も関係しているhas observed · has been produced · has been completed

注意したい発音

話し手はisn't, doesn't, don'tなど、短縮形や弱形を9回使っています。聞こえたとおりの短い形で発音しましょう。

  • 「sh」と「zh」の音: exception /əkˈsɛpʃən/, measurable /ˈmɛʒəɹəbəl/
  • 長い単語(アクセントの位置に注意): interactive /ˌɪn.tɚˈæk.tɪv/, autonomy /ɔˈtɑ.nə.mi/, autonomous /ɔˈtɑnəməs/, measurable /ˈmɛʒəɹəbəl/

日本語話者が苦手な音:

  • /v/ — /b/ にならないように、上の歯を下唇に当てる: interactive /ˌɪn.tɚˈæk.tɪv/, observe /əbˈzɜːv/, involvement /ɪnˈvɑlv.mənt/, approve /əˈpɹuːv/, avenge /əˈvɛnd͡ʒ/
  • /f/ — 「フ」ではなく、上の歯と下唇で出す: differ /ˈdɪfɚ/, specify /ˈspɛs.əˌfaɪ/, fuzzy /ˈfʌzi/
  • 子音の連続 — 間に母音を入れない: upgrade /əpˈɡɹeɪd/, clock /klɑk/, adjust /əˈd͡ʒʌst/, prompt /pɹɑmpt/, introduce /ˌɪntɹəˈdus/

この動画での練習方法

  1. まず声を出さずに動画を最後まで聞き、知らない単語をメモします。
  2. 通常の速度で一文ずつシャドーイングし、話し手のリズムに合うまで繰り返します。
  3. 自分の声を録音して元の音声と比べます。loop, delegate, clockなどの単語に特に注意しましょう。

シャドーイングとは?英語上達に効果的な理由

シャドーイング(Shadowing)は、もともとプロの通訳者養成プログラムで開発された言語学習法で、多言語習得者として知られるDr. Alexander Arguelles によって広く普及されました。方法はシンプルですが非常に効果的:ネイティブスピーカーの英語を聞きながら、1〜2秒の遅延で声に出してすぐに繰り返す——まるで「影(shadow)」のように話者を追いかけます。文法ドリルや受動的なリスニングと異なり、シャドーイングは脳と口の筋肉が同時にリアルタイムで英語を処理・再現することを強制します。研究により、発音精度、抑揚、リズム、連音、リスニング力、そして会話の流暢さが大幅に向上することが確認されています。IELTSスピーキング対策や自然な英語コミュニケーションを目指す方に特におすすめです。

シャドーイングのやり方: ステップ別の完全ガイドを読む →