Big Idea

The argument in one line.

Codex's real power shows up once you go past a single chat window — picking the right model/effort tier per task, delegating work to spawned sub-threads, running unattended goal-loops, and hard-blocking destructive commands with a hook are what separate casual use from a thousand-hour power user.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You already use Codex, Claude Code, or a similar CLI coding agent daily and want concrete settings, not a beginner install guide.
You run more than one agent thread or task at a time and want a way to delegate and supervise them instead of babysitting one chat.
You maintain an AGENTS.md or CLAUDE.md file and haven't audited it since your last model upgrade.
You give an AI agent full filesystem access and want a concrete example of a hook that blocks destructive commands before they run.

SKIP IF…

You've never used an AI coding CLI — this assumes you already know what a thread, a model picker, and a plugin/MCP are.
You want an explanation of how Codex or GPT-5.6 works internally — this is a settings-and-workflow tour, not a technical deep dive.

TL;DR

The full version, fast.

After 1,000+ hours in Codex, the presenter's real habits go past picking a model: for hard problems use the largest tier (Sol) at high effort, for everything else use the smallest (Luna) at high/extra-high effort, since mid-tier Terra scores worse than it costs. Threads can spawn and supervise other threads with their own model/effort, which turns one chat into a small team. AGENTS.md needs a stale-rule cleanup pass after every model release. Goal-loops let an agent run for hours against a private benchmark until a pass-rate target is hit, no re-prompting required. After a viral post about GPT-5.6-Sol deleting a user's files, the video walks through adding a PreToolUse hook that hard-blocks destructive shell prefixes (rm -rf /, deleting $HOME, disk erase commands) before they execute, plus setting approvals to 'approve for me' instead of full unrestricted access.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:36

01 · Intro

Credibility hook (1,000+ hours in Codex) and a subscribe ask before the tips start.

00:36 – 02:49

02 · Model selection

GPT-5.6's Sol/Terra/Luna sizes plotted on an intelligence-vs-cost chart; Terra scores worse than Luna at higher effort, so the rule of thumb is Sol-high for hard problems, Luna-high/extra-high for everything else.

02:49 – 04:54

03 · Threads

Threads can see and spawn other threads; a rule in AGENTS.md can auto-spin a new thread on a specific model whenever the operator says 'deploy,' and one thread can supervise several running in parallel.

04:54 – 05:38

04 · Agents.md review

Asking Codex to review AGENTS.md for stale rules after a new model release, then telling it to clean the file up.

05:38 – 06:47

05 · Sponsor: Zapier

Mid-roll sponsor segment for Zapier's MCP, which connects Codex to 9,000+ third-party apps (Gmail, Trello, Asana, Google Docs, etc.).

06:47 – 08:32

06 · Browser use

Codex's built-in browser can import cookies/passwords and perform real-world tasks; demoed by moving two files into a new archive folder in Google Drive.

08:32 – 09:12

07 · Skills

Skills extend what Codex can do; demoed by installing Matt Pocock's public skills repo via a pasted GitHub URL.

09:12 – 11:28

08 · Loops

Goals/loops let Codex work unattended toward a target for hours; a Loop Library and a 'Loopy' skill help find or draft loops, demoed with a benchmark-improvement loop targeting a 90% pass rate.

11:28 – 13:00

09 · Remote

Pairing a phone with a desktop Codex session via QR code to view and control running threads remotely while the work still executes locally.

13:00 – 14:58

10 · Hooks

A real incident (GPT-5.6-Sol deleting a user's Mac files) motivates adding a PreToolUse hook with explicit forbidden command prefixes (root/home deletion, disk erase/partition) as a hard safety backstop.

14:58 – 15:48

11 · Access

Approval settings compared — full/unrestricted access (what the presenter personally uses but doesn't recommend) versus 'approve for me,' which only pauses on commands flagged as potentially severe.

Atomic Insights

Lines worth screenshotting.

In the GPT-5.6 family, the mid-tier model (Terra) scores lower on the intelligence-vs-cost chart than the smaller model (Luna) at higher effort settings — meaning the 'medium' option is not the safe default it looks like.
The presenter's working rule: the largest model at high effort for hard problems, the smallest model at high or extra-high effort for everything else — he says he almost never uses the largest model at its lightest effort setting.
Running a model in 'fast' mode costs 2.5x more for only 1.5x the speed increase, so it's a bad trade unless you're actually time-constrained.
Threads in Codex can see and spawn other threads, so a main thread can kick off sub-threads on a different model/effort tier and supervise their progress instead of managing each one by hand.
A rule can be written directly into AGENTS.md so that any time the operator says 'deploy,' Codex automatically spins up a separate project-scoped thread using a specific model and thinking level to handle it.
AGENTS.md files accumulate stale rules from prior model generations; the habit worth copying is re-auditing that file every time a new model version ships, since old rules can silently misdirect a newer model.
Goal-loops (Codex calls them 'goals') can run for hours — sometimes up to twelve hours straight — against a private benchmark, iterating on failures and re-running the benchmark until a target score like 90% is hit, with zero re-prompting.
A Codex desktop session can be remote-controlled from a phone via QR-code pairing, and the connection still runs the work locally on the paired desktop rather than in the cloud.
After a widely shared post about GPT-5.6-Sol running an incorrect $HOME-expanding cleanup command and deleting a user's files, the fix demonstrated is a PreToolUse hook that inspects the exact shell command before execution and hard-blocks specific destructive prefixes (rm -rf /, rm -rf ~, disk erase/partition commands) with a stated justification per rule.
The presenter admits he personally runs Codex on full, unrestricted access and says he would not recommend that setting to most people — the safer default he recommends is 'approve for me,' which only pauses for commands the model itself flags as potentially severe.
Skills (also called plugins in some tools) are described as still underrated even though widely known; installing one is as simple as pasting a GitHub URL into a 'install this skill' prompt.

Takeaway

Power use of an AI coding agent is a workflow, not a model choice.

WHAT TO LEARN

The gap between casual and power use of an AI coding agent isn't which model you pick — it's whether you delegate to sub-threads, audit your instruction file, run unattended goal-loops, and hard-block destructive commands before they can run.

02Model selection

In a tiered model family, plot intelligence against cost before assuming the mid-tier option is the safe default — here the mid-tier model scored worse than the small model at higher effort.
A simple two-tier rule of thumb (largest model + high effort for hard problems, smallest model + high/extra-high effort for everything else) covers most day-to-day tasks without per-task decision fatigue.
Running a model in 'fast' mode is usually a bad trade: 2.5x the cost for only 1.5x the speed.

03Threads

An agent thread that can see and spawn other threads turns a single chat into a small delegated team, with each sub-thread able to run on a different model/effort tier.
A standing rule in the instruction file (e.g. 'whenever I say deploy, spin up a separate thread on model X') automates a repeated workflow instead of re-typing it each time.
One thread can be dedicated purely to supervising several others running in parallel, re-prompting any that stall instead of a human checking each one manually.

04Agents.md review

A persistent instruction file (AGENTS.md, CLAUDE.md) needs a stale-rule audit every time the underlying model changes — rules written for an older model can silently misdirect a newer one.
Reviewing then explicitly asking the agent to 'clean it up' is a two-step habit: first surface what's stale, then let the agent do the edit.

06Browser use

Zapier's MCP integration exposes thousands of third-party apps to an agent through one connection, rather than building a custom integration per app.
A browser-capable agent can be pointed at cookie/password import and then given real-world tasks (file organization, DNS changes, resizing a cloud database instance) beyond just testing a website.

07Skills

Installing a shared skill is as simple as pasting its repo URL into an 'install this skill' instruction — the setup friction that made skills feel niche is mostly gone.

08Loops

A goal/loop lets an agent iterate against a private benchmark for hours (sometimes up to twelve) without re-prompting, fixing failures and re-running the benchmark until a target score is hit.
A public loop library and a purpose-built skill for drafting loops both lower the barrier to writing a working goal-loop from scratch.

09Remote

A phone can pair with a desktop agent session via QR code and view/control it remotely, while the actual work still executes locally on the paired machine.

10Hooks

A real, current incident (a model deleting a user's files via a mis-expanded cleanup command) is the concrete justification for treating agent safety as a config problem, not a trust problem — you can't stop a model from writing dangerous code, but you can stop it from running.
A PreToolUse hook that inspects the literal command before execution and blocks specific destructive prefixes (root delete, home directory delete, disk erase/partition) is presented as the actual fix, with each rule carrying its own stated justification.

11Access

The presenter discloses his own default is full, unrestricted access and explicitly says he wouldn't recommend that to most people — a useful signal that convenience and safety trade off even for a heavy daily user.
'Approve for me' is framed as the practical middle ground: low-risk commands run without interruption, but anything the model itself flags as potentially catastrophic still stops for a human decision.

Glossary

Terms worth knowing.

Thread: A single task/conversation session inside Codex; threads can see and spawn other threads, allowing one thread to delegate work to and supervise several others.
Thinking effort: A per-request setting (Light/Medium/High/Extra High/Ultra) that controls how much reasoning budget a Codex model spends on a task — higher effort costs more and consumes usage limits faster.
Goal / loop: Codex's term for an open-ended instruction with a stopping condition (e.g. 'keep iterating until benchmark score reaches 90%') that lets an agent work unattended for extended periods instead of stopping after one response.
AGENTS.md: A markdown file of persistent instructions and rules loaded into every Codex agent's context window for a project — the equivalent of Claude Code's CLAUDE.md.
PreToolUse hook: A hook that runs before Codex executes a tool call (like a shell command), letting a script inspect the exact command and block it if it matches a forbidden pattern.
MCP (Model Context Protocol): A protocol that lets an AI agent connect to external tools and services (in this video, Zapier's MCP, which exposes thousands of third-party app integrations to Codex).
Skill: A packaged, installable set of instructions or capabilities added to an agent (Codex, Claude Code, Cursor, etc.) via a shared URL, extending what the agent knows how to do.

Resources

Things they pointed at.

05:38toolZapier MCP ↗

08:48toolMatt Pocock's skills repo

09:37toolLoop Library ↗

10:01toolLoopy skill ↗

13:16linkDestructive-command hook prompt (GitHub gist) ↗

Quotables

Lines you could clip.

02:20

“For your hardest problems, go with Soul. For everything else, go with Luna.”

crisp, standalone rule of thumb with no setup needed→ TikTok hook↗ Tweet quote

10:24

“You literally say, here's the overall goal. I'm not even gonna tell you how to solve it. Just go solve it and keep working until you do.”

punchy framing of autonomous agent loops vs. single-prompt chats→ IG reel cold open↗ Tweet quote

13:10

“GPT five point six Soul just accidentally deleted almost all of my Max files, and this is why I trust Fable a thousand times more.”

shocking, high-stakes quote (read from a viral post) that sets up the whole hooks section→ TikTok hook↗ Tweet quote

15:02

“I usually leave approvals as full access, unrestricted access to the Internet, any file on your computer, and I actually would not recommend doing that for most people.”

candid, contrarian self-aware admission — presenter flags his own risky default→ newsletter pull-quote↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphor

There are so many unknown tips and tricks to get the most out of Codecs, and even advanced users oftentimes don't know all of them.

That's what this video is about. I put together everything I have learned about Codex after using it for over a thousand hours.

And it doesn't matter if you're a beginner or you're very experienced, you will get something out of this video. And I talk a lot about how to get the most out of artificial intelligence on this channel. And if you appreciate that, please drop a like, hit the subscribe button.

It does help the channel. Thank you in advance. Now that GPT 5.6 is out, you have so many different options to choose from even just within the 5.6 family, whether you're choosing Sol, Terra, or Luna, or the five levels of thinking effort or fast versus regular.

And so knowing when to use which model, which thinking effort is an entire skill in itself. Now I don't wanna get too technical here, but I did wanna show this graph really quick just to show you the differences between SOL, Terra, and Luna, which are the three sizes of the GPT 5.6 model family.

Luna's the smallest, Terra is in the middle, and Sol is the largest. And they should be relatively as capable as their size, but it turns out it's not actually that way.

So on the y axis, what you're seeing is the total score. So higher equals better. On the x axis, what you're seeing is the budget.

So to the left is actually cheaper. To the right is more expensive. And so where you wanna be is as high up as possible and to the left as possible.

So you really wanna be in this quadrant. And so what you're seeing is SOLE definitely scores the highest. Terra has a lower score pretty much than all of the relevant Luna models, the smaller ones.

So if you're on LunaMax, you are paying less and getting more performance out of the model than if you were going to use Terra high.

All of this, I know sounds a little bit confusing. Just think about this. For your hardest problems, go with Soul.

For everything else, go with Luna. So generally, what I'm using is Soul. And for the effort, I'm usually sitting at the high effort level.

Sometimes I'll go down to medium. I basically never use Soul Light. Then for everything else, for the easier task, I'm going with 5.6 Luna, and I'm doing high or extra high on Luna.

Now whether or not you use fast is up to you. If you have time to spare, I recommend not using fast because it is two and a half times as expensive, and you only get one and a half x the speed up. And by the way, there are a number of ways to mix and match the models so you're using the right model for the right task.

So for example, in the agents dot m d file, I can write a rule that says anytime I do some kind of work, I automatically want it to spin up a new thread in Codex using a certain model size and thinking effort and kick it off there.

And that's actually really cool. I'm gonna come back to that in a second. But, basically, what it is is whenever the operator says deploy, and that just means take the code that I've written and put it on my website or my app, immediately create a separate project scoped thread in the exact same project folder as the original thread.

And here's the key, using model GPT 5.6 Luna with max thinking. And here's something that I don't think most people know. In Codecs, each of these on the left side are called threads.

Each of those threads can see all of the other threads.

And so that allows you to do a few really interesting things, one of which is kicking off a thread from an existing thread. That allows you to delegate work to other models, other thinking budgets, while your main thread is managing it.

So here's just an example of what that might look like. Create a new thread using GPT 5.6 LunaMax and give it the prompt to reply with hello.

Now what you're seeing here is we're using GPT 5.6 SoulLite fast.

Let's just hit enter and let's see what happens. So I'll create the task with that model and send that exact prompt. There we go.

So here's the task. You can just click it right there and reply with hello, sent by Codex from another task, and hello.

So it's a very cool feature that allows you more fine grained control over what's going on in Codex. Another thing this allows you to do is manage a bunch of threads from a main thread. So let's say you have a bunch of threads running in parallel and you just wanna babysit them.

You wanna make sure that they're running and that they're progressing appropriately. Then you can spin up one thread to watch all the other ones. If one of them stops, that master thread will see it and then prompt it again and push it along.

So another thing now that GPT 5.6 is here is to take another look at your agents dot m d file. That file probably has accumulated a lot of rules from previous codex models.

And what happens is a rule that applied back then may not apply anymore. So every time a new model release comes out, this is something that we like to do. Review my agents dot m d file for any stale rules or things we should revise or remove.

Alright. So it actually found a bunch of really stale things in my agents dot m d file, so all I'm gonna say is clean it up. And so just take a look at your agents dot m d file because remember, those get loaded very frequently into the context window of every agent that you're using.

And another thing you should be using are plugins. Plugins are awesome, and they allow you to give a ton of additional functionality to Codecs, including from the sponsor of today's video, Zapier.

Let me show you how to set that up. You're gonna go into settings. You're gonna come down here on the left side under integrations and click plug ins, then click on the MCPs, and you can see I already have this Zapier plugin right there.

And what's really neat about the Zapier MCP is you basically get access to Zapier's over 9,000 different applications. So whether you're using Gmail, Trello, Asana, Google Docs, almost everything, you literally can just plug it right into Codecs.

And I've actually been using Zapier for well over ten years at previous companies of mine, and I am a huge fan of what they're doing. They have been building automations for a very long time, they know exactly what they're doing.

So if you want safety, security, and confidence in knowing that it's going to work the first time, go check out Zapier MCPs. I'm gonna drop a link down below so you can find it easily.

And thank you again to Zapier for sponsoring this video. Alright. The next thing that you have to use is the browser encodecs.

It is incredible. Browser use has gotten so good lately, especially with GPT 5.6 that I foresee the Codex Browser and the Cloud Code Browser, which they just released, becoming my daily driver browsers by the end of this year, if not sooner.

That's how frequently I'm using it. And you can now import all of your cookies and all of your passwords directly into the Codex browser, and it is dead simple.

As soon as you open up the browser, it's gonna say, do you wanna import your cookies and passwords? You just say yes, and it does it. You should be using the Codex Browser for everything, not just testing your website or the app that you're building, which is awesome, but you can literally do anything.

Any real world work you can do. So here's an example. I have this folder, and this I know is overly simplistic, but you could do a lot with this.

And I basically wanted to take these two documents, this Cisco talk and this Hermes agent one, and put it in a new archive folder. So I'm just gonna hit enter. I gave it a short prompt and you're actually gonna watch it work right now.

So I've done some really complex tasks with this including changing DNS records. I've also auto scaled up and down my Supabase instance, and here you go. You can actually see it working right now.

It's not only very accurate, it's actually pretty darn fast now too. It's creating a new folder. Gonna name it archive.

Create. There's that new folder. Now hopefully, it puts those two files in there.

Alright. So there we go. It finished.

So you can really do some awesome complex things with the browser in Codex. Alright.

The next thing I wanna talk about, and you probably have heard of it, you probably also use it, but it's still so underrated, and that is skills. You should be adding a ton of very useful skills to your codex, to Clog code if you use that, to Cursor if you use that.

Skills, I really think, are still very underrated. And this is one of my favorite skills if you're doing any coding. This is Matt Pocock's skills skill, and it has a bunch of very useful skills if you're doing any type of coding.

So to add it, you copy the URL. You come into Codex. You say install this skill, and then just paste it in there, and it'll know what to do.

It'll do the rest for you. Alright. I know I've talked about this a lot, but the next thing I wanna talk about are loops.

In Codecs, they are called goals. It basically tells Codecs to keep working towards some overarching goal, and you find that it works for a very long time, hours, sometimes even days.

And loops are really powerful when you have a specific goal in mind, and that goal can be a number of different things. And if loops sound very foreign, I put together a loops library so you can explore other people's loops that they've been using very successfully. So this is the loop library.

You can find loops from myself, from Jason at OpenAI, from Peter Steinberger from OpenAI, and all of these loops are tried and true and work very well.

I also created the loopy skill, which you can install into Codecs, and it'll help you create loops based on your code base. It'll help you draft loops based on what you're trying to accomplish.

It'll also help you find existing loops. All of this is very simple. I'll drop all of these links down below.

And loops are especially powerful because you can set a goal for them, and the agent will just keep working until it reaches that goal. No more giving a single prompt, waiting thirty minutes, and then it gives you a single answer back. You literally say, here's the overall goal.

I'm not even gonna tell you how to solve it. Just go solve it and keep working until you do. So here's an example loop that I run all the time.

I do slash goal, run our benchmark. When you're done, examine the ways that ASTRO failed, and attempt to fix them. Continue iterating on this until our benchmark score reaches 90% or above.

And so I basically wrote my own custom private benchmark for this project that I'm building. And I'm telling Codex, run that benchmark.

And if the score is under 90%, look for why it's under 90%, fix it, and then run the benchmark again. And sometimes this runs for, you know, up to twelve hours straight just consistently improving over time.

And what I found is, let's say I'm starting at 70% pass rate on the benchmark, it eventually gets all the way up to 90%, and I don't have to do anything else. I just kick off the goal and it goes.

Alright. For those of you who are absolutely addicted to codecs, this one's for you.

I'm surprised by how few people know about this, but here's the thing. You can control your codecs on your computer from your mobile device anywhere in the world, and here's how you do it.

So you're gonna go into settings. You're gonna go here on the left side and click connections. And right here, says control this Mac.

You're gonna click this little add button here. Then on your mobile phone, you're gonna scan this QR code from Codex, and it's going to set up a connection between your mobile phone and your desktop. All of this still runs technically locally.

So you have Codex running on your computer, you have Codex Mobile running on your phone, and you have one controlling the other. Once you have that connection set up, you're going to come over here where it says remote. You're going to click that, and this is actually a live view directly into my desktop's codecs.

So you can see right here. Here's the Astro Hub folder. Here's the Matthews Mac Studio six local, and that little green dot means it's actively connected.

And so I can kick it off from here, and it will control my desktop. Now this is useful if you want to just continue the things that you're working on.

You can also control your computer from anywhere in the world from your phone by doing this because Codecs can control your computer, and now you can control Codecs. So super useful. Highly recommend you set this up, especially if you're like me and you want to code on the go.

Alright. This next one is quite serious. Matt Schumer, friend of the show, tweeted this out a few days ago.

GPT five point six Soul just accidentally deleted almost all of my Max files, and this is why I trust Fable a thousand times more.

Now even though he trusts Fable, I love GPT five point six. So I wanna make sure that this never happens to me. There are a few ways to do that, and this is what we're gonna be talking about in this section, overall security and safety.

And so am will posted about it, and this is really the best way to do it. Block this with a hook.

I would suggest that you do that. Okay. So let me show you what that actually means.

When AI writes code, let's say it says, okay. I'm gonna write code to remove every single file on your desktop. I don't want that to happen, but I can't definitively prevent AI from writing that code.

Sometimes it'll hallucinate. This probably won't happen to you, but there is really no way to prevent it definitely. And so when Codex actually goes to run the code, that's when we can actually prevent it from running.

Here's what you wanna do. And by the way, I'm gonna drop this prompt in a GitHub gist down below so you can just copy paste it. So set up these as real Codex Lifestyle Hooks by following the official documentation.

So what we're gonna do is prevent deletion of the root file system, deletion of contents of the root file system, delete the entire home directory, deleting all user directories, and so on. And so after all of this, hopefully, we will prevent any other catastrophic issues.

But you can even get more fine tuned than that. You can have any rules that you want. This is just what I have.

Now if you go into settings, click hooks over here, so pretool use. And what the tool use is is Codex will call a tool to execute a command on your computer, and we want it to happen.

We want that filter to happen before that happens. So right here, we can see hook one, and it's all configured. And the last thing on safety.

I usually leave approvals as full access, unrestricted access to the Internet, any file on your computer, and I actually would not recommend doing that for most people. Most of all, you want to do this approve for me. It's basically a model that determines which commands should just be run without your approval, and then whenever it gets to a command that it thinks is severe or has the chance to do some kind of catastrophic change on your computer, it's going to ask you for approval.

So I highly recommend most people just do approve for me. So that's it. Those are all the suggestions I have for you for today.

Hopefully, this levels up your codex game, and I actually created a whole another video specifically about vibe coding. And so if you wanna see that, check it out right here.

The Hook

The bait, then the rug-pull.

The pitch is credibility, not novelty: after a thousand-plus hours in Codex, the presenter says even advanced users are missing pieces of it. The video backs that up mid-way through with a real, current incident — a viral post about GPT-5.6-Sol deleting a user's files — that turns the back half from tips into a concrete lesson on locking an agent's shell access down before it runs a command you didn't mean to approve.

Frameworks

Named ideas worth stealing.

02:20concept

Sol / Terra / Luna model-tier rule

Plot any tiered model family's intelligence-vs-cost chart and pick from the top-left quadrant rather than assuming the mid-tier model is the safe default — here Terra scores worse than Luna at higher effort.

Steal forChoosing model/effort combos in any provider with a tiered model family (small/medium/large, or fast/standard).

02:49concept

Thread delegation + supervision

One agent thread spawns others on a different model/effort and periodically checks whether they're still progressing, instead of one person babysitting every task manually.

Steal forAny agent harness that supports sub-agents or parallel threads (Claude Code sub-agents, Codex threads).

04:54concept

AGENTS.md staleness audit

Re-review the persistent instruction file for rules that applied to a prior model generation but misdirect the current one, every time a new model version ships.

Steal forMaintaining CLAUDE.md, AGENTS.md, or any other always-loaded agent instruction file.

13:00concept

Destructive-command PreToolUse hook

A hook script reads the exact shell command before it executes and hard-blocks specific destructive prefixes (rm -rf /, rm -rf ~, rm -rf $HOME, disk erase/partition/format) with a stated justification per rule.

Steal forAny coding agent given real filesystem or shell access — this is a direct swipe target for a Claude Code hook.

14:58concept

Approval mode: 'approve for me'

A middle-ground autonomy setting that runs low-risk commands without asking, but pauses for approval on anything the model itself flags as potentially severe or catastrophic.

Steal forTuning the permission/approval model on any agent CLI that supports graduated autonomy.

CTA Breakdown

How they asked for the click.

VERBAL ASK

05:38product

“go check out Zapier MCPs. I'm gonna drop a link down below so you can find it easily.”

Mid-roll sponsor segment placed right after the two free power-tips (model selection, thread delegation), demonstrated live inside Codex's own plugin/MCP settings rather than cut away as a separate ad break.

MENTIONED ON CAMERA

05:38toolZapier MCP ↗

09:37toolLoop Library ↗

10:01toolLoopy skill ↗

13:16linkDestructive-command hook prompt (GitHub gist) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.