跟读练习: Context Engineering for AI Agents: 4 Failure Modes to Avoid - 通过视频学习英语口语
选择练习模式
每种模式都会打开全屏练习室。
116 句
如果句子过短或过长,请点击 Edit 进行调整。
1
context windows have grown from well a few thousand tokens to well a lot bigger like a million
2
or more and that's room for more stuff more instructions tool results docs you name it
3
so if we think about ai agents
4
and we give them access to this big old context window well
5
they should probably be getting better at long -running tasks shouldn't they?
6
I mean the more context the better job they can do.
7
Well a recent study looked at 18 frontier models
8
and it turned out that they all got less reliable as the context window got bigger.
9
Researchers have taken to giving this thing a name and that is context rot.
10
The more tokens in the window the less reliably the model uses any of them.
11
And this is the problem that context engineering is there to solve.
12
Context engineering is deciding what actually goes into this context window at each step of an agent's run.
13
Now there are many ways an agent's context can go wrong and I'm going to cover four of them.
14
The four failure modes that you can encounter when working with these big context windows.
15
And because I'm feeling nice I'll also suggest how to fix each one.
16
And the funny thing about those fixes is they involve deliberately throwing away information.
17
So how does an agent get better by holding less context?
18
Well to answer that we should first take a look at what's actually filling up these giant context windows.
19
So let's say an AI agent was given the job to migrate an aging code base over to a new database driver.
20
So I'm talking about here we are building a migration AI agent.
21
Now at the beginning of the context window the first thing most likely in here is the system prompt.
22
That's the standing instructions for how the agent operates in general.
23
Now the agent task itself
24
that will also be stored in here as well just as a prompt on its own
25
and it might not be very big it might only be a couple of lines.
26
Then many agent harnesses also pull in a bunch of project instructions.
27
Now these are things like an agents .md file explaining how this particular project is built.
28
Then after that we might have some tool definitions, a description of every tool the agent can call.
29
And from here on in the context window turns into a bit of a running log
30
so we're going to have a message history here this is every step the the agent has taken
31
and after let's say an hour of running well
32
this could get pretty big that's going to be a lot of steps
33
and then the biggest band is probably the tool results
34
that we store here which is things like file contents
35
and test output and error logs and schema dumps
36
so the job description the prompt is a few lines near the front
37
and pretty much everything else is the stuff
38
that the agent kind of piles up along the way oh
39
and buried in this as well there happens to be a note
40
this is a note the agent wrote itself it says
41
that the new driver retries failed connections automatically
42
so the old retry code can be deleted
43
so the note just basically says hey don't bother with writing any retry code why do i mention
44
that we'll come back to
45
that anyway why is a full context window a problem well
46
a model doesn't look things up in its context the way
47
a database looks up a row every token in the window
48
is competing for a finite amount of attention essentially an attention budget.
49
And to me, this is kind of the crux of it.
50
The window might hold a million tokens, but the attention spread across those tokens is finite.
51
Now, what about those failure modes?
52
Let's start with the first one.
53
Number one, that is context poisoning.
54
So I cryptically mentioned the agent wrote itself a note about the new driver retrying failed connections automatically.
55
Well it turns out that that note is completely wrong.
56
The agent shouldn't have done this.
57
It basically skim the driver's documentation, it misread it, and then it's written down a false conclusion.
58
And that's all poisoning is.
59
A mistake gets into this context window somewhere and then gets treated as fact because to the AI model,
60
its own notes, they look like the ground truth.
61
So file after file, our agent has been deleting perfectly good retry code.
62
Next, number two, that is context distraction.
63
This is where the sheer size of the history starts steering the agent.
64
Now a language model at its heart is just a pattern continuation machine
65
and when the context holds at say an hour of an agent's own past actions, that history becomes pretty much the strongest pattern around,
66
stronger than the reasoning it picked up in training.
67
So instead of asking what the job needs next the model
68
might drift into doing more of what it has already done
69
in this message history agents in this state can get stuck just kind of going over old ground
70
what about number three that is context confusion
71
and this one comes from the tools our agent here is wired into quite a lot of stuff
72
there's probably some amount of version control here
73
it might be access to a database console maybe it also is logging tickets into a ticketing system
74
and each of these tool definitions is in the window whether it actually gets used
75
or not and having more of them makes choosing the right tool a bit harder
76
and in one benchmark a model failed a task
77
when it was offered 46 tools but it then passed the very same task
78
when it was only offered 19 tools
79
that is tool sprawl work okay number four that is context clash
80
which is when the window contains information that disagrees with itself
81
so take that note with the bad information at some point our agent is going to see
82
that no retries are happening and it's going to figure out
83
that that note was wrong and it will probably make a note of
84
that into its message history as well
85
but the problem then is we now have both versions sitting in the window
86
that are presenting conflicting information so these are four different ways a context window can go bad
87
all right so what about some potential fixes well for poisoning one of the fixes is to validate.
88
Before a conclusion gets written down into this context window, check it against something real.
89
So one quick connection test and our agent would never have written this no retry code note.
90
What about distraction?
91
Well, the fix there is really too compact.
92
When the history gets really, really long, the agent can summarize it,
93
which is to say it throws the full log away and starts again from the summary,
94
a shortened version of what was in this message history in the window.
95
Now compaction is lossy.
96
A summary of a summary can start to drift like a photocopy of a photocopy.
97
So consider moving as much out of the context window as possible and put them into permanent files instead.
98
Detailed information can be retained in those files, and then it can be pulled into the context window only when it's actually needed.
99
All right, what about context confusion?
100
Well, that is a curation problem, so the fix is to curate.
101
The agent doesn't need 46 tools for database migration, so just give it the handful that the job calls for,
102
and then it can always go and fetch others
103
if it actually needs them then finally context clash well the fix there can be to prune so
104
when information gets superseded just delete the old version rather than leaving both in the window
105
we could have deleted this guy entirely or go further
106
and start a fresh window just restating the current corrected state of the job
107
and leave the poison stuff behind so
108
if a false conclusion gets saved that's poisoning so you can validate
109
if the history has taken over that's distraction so compact
110
if the agent has more tools than the task needs that's confusion
111
so curate and if the context disagrees with itself that's a clash and the fix is prune.
112
And all four fixes work by throwing something away.
113
An agent can get better at long tasks by holding less.
114
Now I'd love to know how you keep an agent's context lean.
115
Do you have your own context engineering fixes?
116
Let me know in the comments.
同频道
FOR YOU
推荐视频
本课的词汇与口语要点
这节口语课以视频“Context Engineering for AI Agents: 4 Failure Modes to Avoid”为素材。 视频中反复出现的词有:context, agent, window, tool, note。 这段视频共有 100 个句子、1451 个单词可供跟读。 讲话部分时长为 10:08。 说话人语速平稳,每分钟约 143 个词,很适合跟读。 86% 的单词属于英语最常用的 3,000 词,其余的词建议在练习前先学一下。
视频中的重点词汇
视频中 15 个值得学习的单词,附发音和释义:
| 单词 | 发音 | 释义 |
|---|---|---|
| poison 名词 | /ˈpɔɪ.zən/ | 毒, 毒物 |
| token 名词 | /ˈtoʊkən/ | 代幣 /代币, 籌碼 /筹码 |
| database 名词 | /ˈdeɪtəˌbeɪs/ | 數據庫 /数据库 |
| delete 动词 | /dɪˈliːt/ | 刪除 /删除 |
| conclusion 名词 | /kənˈkluːʒən/ | 結束 /结束 |
| summary 名词 | /ˈsʌm.ə.ɹi/ | 摘要 |
| confusion 名词 | /kənˈfjuːʒən/ | 混亂 /混乱 |
| instruction 名词 | /ɪnˈstɹʌkʃən/ | 教導 /教导 |
| clash 名词 | /ˈklæʃ/ | 錚錚 /铮铮 |
| log 名词 | /lɔɡ/ | 圓木 /圆木 |
| migration 名词 | /maɪˈɡɹeɪʃ(ə)n/ | 徙動 /徙动, 遷移 /迁移 |
| disagree 动词 | /dɪsəˈɡɹi/ | 不同意, 反對 /反对 |
| compact 形容词 | /kəmˈpækt/ | 緊湊的 /紧凑的 |
| finite 形容词 | /ˈfaɪ.naɪt/ | 有限的 |
| output 名词 | /ˈaʊtpʊt/ | 輸出 /输出 |
视频中出现的短语动词
| 单词 | 发音 | 释义 |
|---|---|---|
| turn out 动词 | 原來 /原来 | |
| figure out 动词 | 弄清楚 | |
| fill up 动词 | 裝滿 /装满 | |
| look up 动词 | /ˌlʊk ˈʌp/ | 仰, 上看 |
| take over 动词 | 接管, 接手 | |
| throw away 动词 | 棄置 /弃置, 扔 | |
| write down 动词 | 寫下來 /写下来, 寫下 /写下 |
视频中的语法
说话人最常用的结构,并附上视频中的原话:
| 结构 | 视频中的用法 |
|---|---|
| 被动语态 be + 过去分词 — 强调发生了什么,而不是谁做的 | was given · be stored · is built |
| 现在完成时 have/has + 过去分词 — 过去发生但与现在仍有关联的事 | have grown · have taken · has taken |
需要注意的发音
说话人用了 12 次缩略和弱读形式,例如 doesn't, gonna, I'm。请按听到的简短形式来说。
- “sh” 和 “zh” 音: conclusion /kənˈkluːʒən/, confusion /kənˈfjuːʒən/, instruction /ɪnˈstɹʌkʃən/, clash /ˈklæʃ/, distraction /dɪsˈtɹæk.ʃən/
- 长单词——注意重音位置: essentially /ɪˈsɛnʃəli/, deliberately /dɪˈlɪb.(ə.)ɹət.li/, documentation /ˌdɑkjəmɛnˈteɪʃən/, continuation /kənˌtɪnjʊˈeɪʃ(ə)n/, photocopy /ˈfəʊtəˌkɒpi/
中文母语者容易读错的音:
- /v/ — 不要读成 /w/,上齿轻触下唇: solve /sɒlv/, involve /ɪnˈvɑlv/, validate /ˈvæl.ɪ.deɪt/
- 词尾辅音 — 要读清楚,后面不要加元音: database /ˈdeɪtəˌbeɪs/, delete /dɪˈliːt/, clash /ˈklæʃ/, prompt /pɹɑmpt/, log /lɔɡ/
- /r/ 和 /l/ 的区别: reliable /ɹɪˈlaɪəbl̩/, deliberately /dɪˈlɪb.(ə.)ɹət.li/, sprawl /spɹɔːl/
如何用这段视频练习
- 先完整听一遍视频,不要开口,记下不认识的单词。
- 用正常速度逐句跟读,每句重复到你的节奏与说话人一致为止。
- 录下自己的声音并与原声对比,特别注意 poison, token, database 这类单词。
什么是跟读法?
跟读法 (Shadowing) 是一种有科学依据的语言学习技巧,最初开发用于专业口译员的培训,并由多语言者Alexander Arguelles博士普及。这个方法简单而强大:您在听英语母语原声的同时立即大声重复——就像是一个延迟1-2秒紧跟说话者的影子。与被动听力或语法练习不同,跟读法强迫您的大脑和口腔肌肉同时处理并模仿真实的讲话模式。研究表明它能显着提高发音准确性,语调,节奏,连读,听力理解和口语流利度——使其成为雅思口语备考和真实英语交流最有效的方法之一。




























