تدريب Shadowing: 900+ hours of Learning AI Agents in 9 Minutes - تعلم التحدث بالإنجليزية عبر الفيديو

اختر وضع التدريب

كل وضع يفتح غرفة تدريب بملء الشاشة.

148 جمل

إذا كانت الجمل قصيرة أو طويلة جدًا، انقر على Edit لتعديلها.

AI agents can already write code, use tools, browse files, and run commands.

So why do they still fail at real engineering tasks?

This is because writing code is only one part of the problem.

The harder part is controlling what the agent does after that first step.

How do you keep one agent reliable?

How do you let it work for hours without supervision?

And how do you coordinate dozens of agents on one massive task?

That's where these three ideas come in.

Harness engineering, loop engineering, and graph engineering.

In this video, I'll break down all three, how they fit together, and when each one actually makes sense.

Hi friends, I'm Maddie.

I'm a senior software engineer who previously worked at Google and interned at other big tech companies like Amazon, IBM, and Microsoft.

In this video, I'll talk you through what harness, loop, and graph engineering mean, how they fit together, and when each is actually worth using.

Let's get into it.

We're going to start with harness engineering.

Harness engineering is the work of designing the agent harness, or everything around the model so the tools it can use,

what goes into its context window, the rules it follows, and the checks it has to pass.

An agent is, simply put, a model plus a harness.

The model is whatever the lab ships, so the harness is the part you actually get to customize.

Here's what that looks like.

Say you're building a budgeting app.

You ask a coding agent to fix a bug where recurring expenses get counted twice,

and it tells you it's done while the tests are still failing.

To fix this, Harness Engineering treats it as a design problem.

You give the agent a tool that runs the test, a rules file that says the task isn't finished until they pass,

and a check that blocks the done message until the test pass.

Now, that mistake can't happen again on this bug or the next one.

Mitchell Hashimoto, the co -founder of HashiCorp, started calling this

Harness Engineering, described simply as this.

Every time your agent makes a mistake, you change the system around it so it can't make that mistake again.

Langchain reported moving their coding agent from around 30th place into

the top five on the Terminal Bench 2 .0 benchmark while keeping the model the exact same.

So all that improvement came from harness engineering.

And in practice, the work falls into four areas you can modify.

Number one, tools, what the agent can do.

Number two, context, what it sees at each step and what you keep out.

Number three, your project's conventions, written down once so it stops guessing.

And number four, feedback.

Tests, linters, and reviewers that tell it when it's wrong.

The tools area includes connections.

An agent can only act on what it can reach, an MCP.

The model context protocol is a standard way a harness plugs into outside tools and data.

And one connection worth adding to your harness is Granola, the sponsor of today's video.

Let's say you just finished a planning call for that budgeting app and you made a bunch of decisions on it.

Normally, your coding agent knows none of them.

But with Granola connected, you can type one line.

So pull the decisions and action items from today's call and turn them into tasks.

And a minute later, you have a task list with exactly what you agreed on.

So for example, your agent now knows that the duplicate expense bug has to be fixed before the export feature ships.

Granola is the AI notepad for people in back -to -back meetings.

It transcribes your computer's audio directly, so nothing extra has join your call.

And it works across

Zoom, Google Meet, Slack, and in -person conversations.

You can optionally also jot a few notes down like you normally would,

and Granola builds a clean summary with the decisions and action items around them.

And what makes that one -line request work is Granola's MCP connector.

You connect Granola to Claude, ChachiPT, or whichever agent you already use, and your meeting context becomes one more thing your harness can reach.

So your agent works from what was actually said in the meeting and you don't have to manually paste anything in.

And with speaker tags on Zoom and Google Meet, you can ask who said what, like what did

Sam say about the launch timing?

If you're a new user, you can get 100 % off your first month through my link in the description.

Okay, now let's get back to the second type of engineering.

So, harness engineering was about one agent doing one task well.

Loop engineering, on the other hand, asks a different question.

Who starts the task?

So, remember our budgeting app.

Right now, nothing happens until you sit down and type the prompt.

You are the trigger.

So every bug fix, every dependency update, every failing test waits for you to notice it.

In June, Adi Azmani on the Cloud Code team at Anthropic published an essay that defined loop engineering as the following.

You replace yourself as the person who prompts the agent, and you design the system that does it instead.

Boris Cherney, who leads

Cloud Code Anthropic, says his job now is to write loops.

And Peter Steinberger, the founder of OpenClaw, said developers should be designing the loops that prompt their agents.

Osmani breaks a loop into five building blocks plus memory.

The first block is automations, which run on a schedule and go find the work.

Then we have work trees, so parallel agents can't overwrite each other's files.

Third, we have skills, which are your project knowledge written down once.

Fourth, we have connectors, which is the MCPP suite just covered.

And fifth, sub -agents.

So the agent that writes the code is never the one verifying it.

And we have a state file on disk because...

Every run starts with a blank memory, so progress has to live somewhere outside the conversation.

For the budgeting app, a loop could look like this.

Every morning, an automation reads last night's failed tests and new bug reports.

For each one worth fixing, it opens an isolated work tree.

One agent drafts a fix, a second agent reviews it gets your test, and a progress file records what passed.

You wake up to PRs waiting for your review.

There are two things I want you to keep in mind here.

First, the hard part of loop is actually the stop condition.

Keep going until every test in the auth folder passes and lint is clean is something a machine can check.

Make the code better is not.

Second, Osmani cautions in the same essay that he's skeptical.

If no one is actively watching the lube, no one is watching its mistakes either.

And the faster it ships code you didn't write, the wider that chasm gets between what exists and what you actually understand.

He calls this comprehension debt.

Finally, let's talk about the newest term, graph engineering.

As you probably learned in a math or data structures class, a graph is nodes connected by edges.

In an agent system, nodes are where work happens, regular code, a single monocle, or an entire agent.

100

Edges decides what runs next, what runs in parallel, and who checks who.

101

Take that budgeting app one more time.

102

Let's say you decide to migrate the whole thing from JavaScript to TypeScript.

103

A single agent working through 200 ,000 order would be slow and might lose track halfway.

104

So instead, you can shape the work as a graph.

105

For example, one node maps the code base and splits it into chunks.

106

That fans out to many agents converting files in parallel.

107

Each result goes to a reviewer node, and anything that fails review travels back along an edge for another attempt,

108

and a final node runs the full test suite.

109

This is the design behind dynamic workflows in Cloud Code.

110

Dynamic workflows have clawed write -in orchestration script

111

that spreads work across tens to hundreds of subagents and checks the results before they reach you.

112

As an example, Jared Sumner used dynamic workflow to port a giant database from Zig to Rust.

113

This was roughly 750 ,000 lines of Rust and took 11 days from first commit to merge.

114

I also want to talk about Langchain's take specifically on graph engineering.

115

Earlier this year, Langchain published a post called Three Years of Graph Engineering.

116

Their argument is that the idea of graph engineering itself is actually old.

117

They've been modeling agents as graphs for about three years.

118

A loop by their account is just a simple graph.

119

But the new part of all this is the node.

120

That node used to hold one model call.

121

Today, it can hold a full agent you trust with real work.

122

Langchain also warns that for open -ended tasks like deep research, where you can't predict the steps ahead of time,

123

they actually moved away from a fixed graph and went back to a single agent running in a good harness.

124

A graph really helps when you know the structure of the work, but when you don't, it just gets in the way.

125

So what should you do with these techniques if you're early in your career?

126

First, don't just memorize these vocabulary buzzwords.

127

Actually understand what's happening.

128

The names of the specific components within the different types of engineering and even the types of engineering themselves, whether it's harness, loop, graph,

129

might keep changing, but a few underlying skills won't.

130

One skill is writing a checkable definition of done, which means tests, evals, and stop conditions.

131

Two is decomposition, deciding what can run in parallel and what depends on what.

132

And three is verification, keeping the maker separate from the checker and reading what your agents produced.

133

So let's put this all together.

134

Harness engineering is designing everything around one model so it's reliable in one task.

135

Loop engineering is designing a system that starts that work and keeps it going until a condition you wrote is true.

136

And graph engineering is wiring many of those together, with rules for who runs when and who checks whom.

137

Each of these builds onto the next, and each one costs more than the previous one.

138

Anthropic published numbers on this when they wrote up how they built Claude's research feature.

139

One agent burns roughly four times the token of a normal chat, and a team of agents burns roughly 15 times.

140

So before you automatically decide to go with the more complex workflow, ask whether the task is valuable enough and parallel enough to pay for it.

141

For example, a bug fix might need a harness, nightly maintenance might need a loop, and a thousand file migration might need a graph.

142

So in conclusion, harness engineering makes one agent reliable, loop engineering keeps that agent running without you, and graph engineering coordinates many of them.

143

They are different layers of the agentic build system, and the skills that carry across all three layers are defining done,

144

breaking work down, and verifying the results.

145

And as As a reminder, if you want your meeting context plugged into your own agent setup, check out Granola through my link in the description.

146

And that's all I have for you in this video.

147

If this was helpful, hit that like button, hype the video, and subscribe for more tech and AI content.

148

Thanks for watching, I'll see you next one!

تنزيل التطبيق

كل ما تحتاجه للتحدث بطلاقة

تقييم النطق بالذكاء الاصطناعيتقييم كل جملة

تدريب IPAأتقن كل صوت

المفرداتابنِ رصيدك من الكلمات

لعبة المفرداتتعلّم أثناء اللعب

امسح رمز QR

Download on theApp Store Get it onGoogle Play

نفس القناة

How I use AI as a senior software engineer and 5 principles of AI explained in 10 minutes

Daily ConversationWorking Adults

The AI Take Over Has Completely Backfired (Here's the Proof)

Business EnglishUniversity StudentsWorking Adults

FOR YOU

فيديو موصى به

26 جمل

How does AI actually work?

Daily ConversationUniversity StudentsWorking Adults

207 جمل

Your AI Agent Starts Every Session Like a New Hire - Killian Carlsen-Phelan, Sonar

Daily ConversationUniversity StudentsWorking Adults

272 جمل

المفردات وملاحظات التحدث لهذا الدرس

هذا الدرس في المحادثة مبني على الفيديو “900+ hours of Learning AI Agents in 9 Minutes”. أكثر الكلمات تكرارًا في الفيديو: agent, engineering, harness, graph, loop. يحتوي هذا الفيديو على 148 جملة و1928 كلمة للتدرب على الترديد. مدة الكلام 9:54. يتكلم المتحدث بسرعة، نحو 195 كلمة في الدقيقة، لذا ستسمع كثيرًا من وصل الكلمات واختزال الأصوات. 81% فقط من الكلمات ضمن أكثر 3,000 كلمة شيوعًا في الإنجليزية، فالمفردات صعبة نسبيًا.

أهم المفردات في هذا الفيديو

15 كلمة من الفيديو تستحق التعلّم، مع النطق والمعنى:

الكلمة	النطق	المعنى
graph اسم	/ɡɹæf/	مُخَطَّط بَيَانِيّ
loop اسم	/luːp/	عُرْوَة, عُقْدَة
parallel صفة	/ˈpæ.ɹəˌlɛl/	مُوَازٍ
connect فعل	/kəˈnɛkt/	رَبَطَ
reliable صفة	/ɹɪˈlaɪəbl̩/	مَوْثُوق
founder اسم	/ˈfaʊ̯n.dəː/	مُؤَسِّس, مُؤَسِّسَة
layer اسم	/ˈleɪ̯ɚ/	طَبَقَة
essay اسم	/ˈɛs.eɪ/	مَقَالَة, بَحْث
expense اسم	/ɪkˈspɛns/	صَرْف, مَصْرُوف
suite اسم	/swiːt/	حَشَم
plug اسم	/plʌɡ/	قَابِس, فِيشَة
verify فعل	/ˈvɛɹɪfaɪ/	أَكَّدَ
automation اسم	/ˌɔ.təˈmeɪ.ʃən/	أَتْمَتَة, تَأْلِيَّة
audio صفة	/ˈɔ.di.oʊ/	سَمْعِيّ, أُودْيُو
technique اسم	/tɛkˈniːk/	أُسْلُوب, تِقْنِيَة

الأفعال المركبة التي ستسمعها

الكلمة	المعنى
sit down فعل	جَلَسَ, قَعَدَ

جمل تستحق التكرار

جمل قصيرة وكاملة من الفيديو يمكنك استخدامها في الحديث اليومي:

How do you keep one agent reliable?
How do you let it work for hours without supervision?
We're going to start with harness engineering.
Say you're building a budgeting app.
Finally, let's talk about the newest term, graph engineering.

القواعد في هذا الفيديو

أكثر التراكيب التي يستخدمها المتحدث، مع العبارات نفسها من الفيديو:

التركيب	في الفيديو
جمل الوصل who / which + جملة — معلومة إضافية عن شخص أو شيء	person who prompts · skills, which are · connectors, which is
المبني للمجهول be + اسم المفعول — التركيز على ما يحدث لا على من يفعله	be fixed · was actually said

نطق يحتاج إلى انتباه

يستخدم المتحدث 20 من الصيغ المختصرة، مثل can't, don't, I'll. انطقها بصيغتها القصيرة كما تسمعها.

أصوات “th”: mathematics /mæθ(.ə)ˈmæt.ɪks/, authorization /ˌɔːθəɹaɪˈzeɪʃən/
أصوات “sh” و“zh”: automation /ˌɔ.təˈmeɪ.ʃən/, conclusion /kənˈkluːʒən/, migration /maɪˈɡɹeɪʃ(ə)n/, caution /ˈkɔːʃ(ə)n/, supervision /ˌs(j)uːpəˈvɪʒən/
الكلمات الطويلة — انتبه لموضع النبر: coordinate /koʊˈɔɹ.dəˌneɪt/, automation /ˌɔ.təˈmeɪ.ʃən/, developer /dɪˈvɛləpɚ/, mathematics /mæθ(.ə)ˈmæt.ɪks/, supervision /ˌs(j)uːpəˈvɪʒən/

أصوات يجدها متحدثو العربية صعبة:

/p/ — ليس /b/: انطقه مع نفخة هواء: loop /luːp/, parallel /ˈpæ.ɹəˌlɛl/, prompt /pɹɑmpt/, expense /ɪkˈspɛns/, plug /plʌɡ/
/v/ — ليس /f/: مع اهتزاز الأوتار الصوتية: verify /ˈvɛɹɪfaɪ/, actively /ˈæktɪvli/, developer /dɪˈvɛləpɚ/, convert /kənˈvɝt/, supervision /ˌs(j)uːpəˈvɪʒən/
تتابع الصوامت — لا تضف حركة بينها: graph /ɡɹæf/, connect /kəˈnɛkt/, prompt /pɹɑmpt/, expense /ɪkˈspɛns/, suite /swiːt/

كيف تتدرب بهذا الفيديو

استمع إلى الفيديو كاملًا مرة واحدة دون أن تتكلم، ودوّن الكلمات التي لا تعرفها.
ابدأ بسرعة 0.75×، وردّد جملة جملة، ثم عُد إلى السرعة العادية عندما يصبح الأمر سهلًا.
سجّل صوتك وقارنه بالأصل، مع الانتباه إلى كلمات مثل graph, loop, parallel.

ما هي تقنية التظليل الصوتي؟

التظليل الصوتي (Shadowing) تقنية تعلم لغة مدعومة علمياً، طُورت أصلاً لتدريب المترجمين الفوريين المحترفين. الطريقة بسيطة لكنها قوية: تستمع لصوت إنجليزي أصلي وتكرره فوراً بصوت عالٍ — كظل يتبع المتحدث بتأخير 1-2 ثانية. تُظهر الأبحاث تحسناً كبيراً في دقة النطق والتنغيم والإيقاع وربط الأصوات والاستماع والطلاقة.

تقنية الشادوينغ: اقرأ الدليل الكامل خطوة بخطوة →

تدريب Shadowing: 900+ hours of Learning AI Agents in 9 Minutes - تعلم التحدث بالإنجليزية عبر الفيديو

اختر وضع التدريب

تنزيل التطبيق

How I use AI as a senior software engineer and 5 principles of AI explained in 10 minutes

The AI Take Over Has Completely Backfired (Here's the Proof)

How does AI actually work?

Your AI Agent Starts Every Session Like a New Hire - Killian Carlsen-Phelan, Sonar

How I'd Learn n8n if I had to Start Over in 2026

GitHub Copilot now controls your command line...

How to Build an AI Agent for Every Part of Your Life

It’s not just you

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

A Practical Guide To Becoming An AI Engineer (2026)

The 4-Hour Workweek in a World of AI

Harnessing generative AI in architecture, engineering & construction

I Tested The Viral Reverse Engineering Repo

Spec-Driven Development: AI Assisted Coding Explained