16:14Riley Brown · Essay
Browsers Are Dead. Codex Just Replaced Them.
A 16-minute breakdown of why AI browsers lost before they launched and how Codex and Claude Code absorbed the browser entirely.
May 28thThe argument in one line.
Prompt hacks and tool mastery are perishable — the only durable edge in the AI agent era is the ability to describe what you want clearly, delegate effectively, and recognize quality output.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code, Codex, Cursor, or any AI agent platform daily and want to understand where the category is heading over the next 12–18 months.
- You're spending time on prompt engineering tricks and wondering whether those skills will age well.
- You want to evaluate cloud agents (Chorus, Poke) vs. desktop super-apps (Codex, Claude) and understand the actual use-case split.
- You're a solo builder or small team trying to decide which AI agent platform to go deep on.
- You're curious about GLM 5.2 and open-source models as cost alternatives to frontier models like Opus 4.8 or Fable 5.
SKIP IF…
- You need deep technical implementation — this is a trends overview, not a how-to guide.
- You're already deeply familiar with Codex, Chorus, and computer-use tools and follow the AI space daily.
- You want rigorous benchmarks or independent cost analysis — the numbers cited are anecdotal from the presenter's own usage.
The full version, fast.
AI models are getting good enough that precise prompting matters more than prompt hacks, and skills (reusable instruction files) are becoming the primary unit of agent productivity. The platform layer is consolidating into super-apps (Codex, Claude Desktop) for computer-side work and cloud agents (Chorus) for asynchronous tasks in Slack, iMessage, and WhatsApp. As agents become ambient — always on, scheduling their own tasks, controlling your desktop — the skills that rise to the top aren't tool-specific: they're communication clarity, delegation instinct, and knowing what good output looks like in your domain. Frontier model costs are rising sharply (one 9-prompt session cost ~$250 via API), while open-source models like GLM 5.2 are closing the gap fast, making OpenRouter-style model switching a practical strategy.
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 →Where the time goes.
00:00 – 00:50
01 · Cold open — the promise
Sets up the frame: this is about nine inevitable, non-perishable trends, not tips and tricks that will be obsolete in a year.
00:50 – 04:50
02 · Trend 1 — Models are getting smarter (prompts go plain)
The 'I want you to act as' hack is dead. Context management via @-mentions in Cursor is fading. GPT Image 2 needs only English descriptions. The enduring skill: describe what you want precisely.
04:50 – 09:05
03 · Trend 2 — Skills are taking over (and will self-assemble)
Skills = reusable instruction files. Don't write them manually — build by doing, then ask the agent to save as a skill. Demo: hook-outline skill created by telling Codex to watch a YouTube video, then turn the result into a skill. Skills auto-update when you give feedback.
09:05 – 13:50
04 · Trend 3 — Agent platforms become super-apps
Codex and Claude Desktop are the two super-apps. They combine chat, skills, automations, in-app browser, coding (terminal), and now site hosting. Multitask by running parallel agents simultaneously. Key differentiator: built-in full browser inside the app.
13:50 – 18:10
05 · Trend 4 — Agents connect to tools we already use
Cloud agents (Chorus) live in iMessage, Slack, WhatsApp, Telegram — never turn off, great for small teams. Demo: request research + landing page through iMessage, get a public URL back. Slack demo: at-mention the agent, it reacts with an emoji, delivers a Hormozi/Theo analysis page.
18:10 – 20:30
06 · Trend 5 — Foundational skills rise to the top
The pyramid: Hacks (top, perishable) → Tools (middle) → Foundations (base, durable). Five foundations: (1) communication, (2) delegation, (3) understanding what good looks like in your domain, (4) mental clarity, (5) multitasking. If you're a good manager of people, you'll be a good manager of agents.
20:30 – 24:30
07 · Trend 6 — Agents work asynchronously (automations)
Any agent task can become a scheduled automation just by asking in natural language. Demo: 'send me a hook outline every morning at 9am' creates a cron job inside Codex. This replaces Zapier-style workflow builders with plain-language scheduling.
24:30 – 27:45
08 · Trend 7 — Agents control your computer
Computer use is built into Codex — click, type, navigate apps autonomously. Three modes: computer use, in-app browser, external Chrome control. Prediction: AI will exceed human computer-control ability within 12–18 months.
27:45 – 32:45
09 · Trend 8 — Frontier costs rise; open-source catches up
Fable 5: 9 prompts = ~$250 API cost. Frontier pricing shows no sign of dropping. Counterweight: GLM 5.2 is near Opus 4.8 quality at far lower cost. Practical move: use OpenRouter to route tasks to the cheapest capable model. Demo: setting up GLM 5.2 in Cursor via OpenRouter API key.
32:45 – 33:39
10 · Trend 9 — Real-time voice + computer control
Built 'Jarvis' desktop app in ~3 prompts on Codex 5.5. Live demo: voice commands open the Comet browser, navigate to LinkedIn, open Cursor, type a prompt, switch to Notion — all via spoken instructions. Thinking Machines also working on this. The worst it'll ever be is right now.
Lines worth screenshotting.
- The 'I want you to act as' prompt hack that dominated 2023 Twitter threads is already dead — models are smart enough that plain English outperforms any framing trick.
- Skills (reusable instruction files) are the primary productivity unit in agent platforms — 100+ skills built iteratively beats any single clever prompt.
- You don't manually write skills — you build them by telling the agent to do a thing well, then asking it to save that as a skill.
- AI agent platforms are consolidating into super-apps: Codex and Claude Desktop are the two main ones, and you should go deep on one.
- Cloud agents (Chorus, Poke) and super-apps serve different use cases — cloud agents live in Slack/iMessage and never turn off; super-apps run on your machine.
- If you're a good manager of people, you'll be a good manager of agents — the delegation skills transfer almost directly.
- You can only delegate effectively to an AI agent in domains where you know what good looks like — domain expertise doesn't get replaced by AI, it gets amplified.
- Frontier model costs are rising with no sign of stopping — a 9-prompt Fable 5 session ran ~$250 worth of API compute.
- GLM 5.2 is reportedly close to Opus 4.8 across the board at a fraction of the cost — open-source Chinese models are the cost arbitrage play right now.
- OpenRouter is the practical way to test open-source models without juggling multiple API keys or provider accounts.
- Automations are just agent tasks with a trigger — anything you do in Codex or Claude, you can schedule with natural language ('do this every morning at 9am').
- Computer use is still slow but improving fast — the prediction is AI will outperform humans at computer control within 12–18 months.
- Real-time voice + computer use is the Jarvis moment — already functional enough to open apps, type prompts, and navigate between tools by voice.
- Context management (manually @-mentioning files in Cursor) is becoming obsolete the same way prompt hacks did — models now infer context automatically.
- AI agents entering tools you already use (Slack, iMessage) is more disruptive than new standalone agent apps — the interface disappears into existing workflows.
Nine trends that won't change, whatever tool wins.
WHAT TO LEARN
The AI agent landscape is shifting fast at the tool layer, but the underlying structural trends are stable — and the builders who orient around those trends will outcompete those chasing hacks.
02Trend 1 — Models are getting smarter
- Prompt hacks ('act as a Linux terminal', 'act as a doctor') are already obsolete — the model quality jump since 2023 makes plain-language descriptions outperform any framing trick.
- This applies across text, code, and image generation — GPT Image 2 needs only English descriptions; Midjourney parameters are no longer necessary.
03Trend 2 — Skills are taking over
- Skills are the primary productivity unit in agent platforms — don't write them upfront; build them by refining real work, then ask the agent to save the result as a named skill.
- When a skill produces a subpar output, just ask the agent to update it — the skill file changes automatically, and future runs inherit the improvement.
- The long-term direction is agents that auto-assemble skills from observed behavior, with no manual creation required.
04Trend 3 — Super-apps
- Codex and Claude Desktop are the two super-apps to know — they combine chat, skills, automations, in-app browser, and terminal in one interface.
- Parallel agent threads (Command+N) let you multitask across several long-running tasks simultaneously inside one platform.
05Trend 4 — Agents in existing tools
- Cloud agents (Chorus) live in iMessage and Slack and never turn off — they're better suited for small-team workflows where you want async responses without opening a desktop app.
- Messaging an AI agent in Slack or iMessage and getting a public URL back as a response is already functional today, not a future state.
06Trend 5 — Foundational skills
- The Foundations Pyramid puts communication, delegation, domain expertise, mental clarity, and multitasking at the base — these are the skills that compound as tools commoditize.
- Domain expertise is the non-negotiable multiplier: if you don't know what a good deliverable looks like, you can't effectively direct an agent to produce one.
07Trend 6 — Async automations
- Any agent task can be scheduled with a natural-language instruction — 'do this every morning at 9am' creates a cron job without any Zapier-style configuration.
- Useful automations to set up immediately: daily research briefs, recurring hook-outline drafts, scheduled content research on competitors.
08Trend 7 — Computer use
- Computer use already works well enough to replace humans on repetitive UI tasks — the gap is speed, not capability, and speed is improving weekly.
- Three access modes exist today: full computer use (screen + clicks), in-app browser (inside Codex), and external Chrome control — each has different latency and reliability tradeoffs.
09Trend 8 — Cost divergence
- Frontier model API costs are not trending down — 9 prompts with Fable 5 ran ~$250 in compute, and enterprise demand is keeping prices high.
- GLM 5.2 from Zhipu AI is reportedly near Opus 4.8 quality at a fraction of the cost — validating open-source Chinese models as a legitimate cost arbitrage for non-critical tasks.
- OpenRouter is the practical tool for model-switching — one API key, all models, zero vendor lock-in.
10Trend 9 — Real-time voice + computer control
- Voice-controlled computer use is already functional — opening apps, typing prompts, switching between tools by speaking is a working workflow today, not a demo.
- The current version is the worst it will ever be — building voice-first agent habits now means compounding returns as the technology improves.
Terms worth knowing.
- Skills
- Reusable task-specific instruction files stored in an agent platform. When you trigger a skill by name, the agent follows those instructions rather than interpreting your prompt from scratch. Equivalent to saved workflows or macros, but written in natural language.
- Super-app
- An all-in-one agent platform that combines chat, coding, browsing, and automation in a single desktop interface. The presenter cites Codex and Claude Desktop as the two primary examples in 2026.
- Cloud agent
- An AI agent that runs on a remote server rather than your local machine, allowing it to stay active 24/7, respond to messages in Slack or iMessage, and run scheduled tasks without your computer being open.
- Computer use
- The ability of an AI agent to visually perceive a computer screen and take actions — clicking, typing, navigating apps — as if it were a human operator. Distinct from API-based tool use, which connects to software back-ends directly.
- OpenRouter
- An API aggregator that gives access to dozens of AI models (OpenAI, Anthropic, open-source) through a single API key, making it practical to swap models per task without managing multiple provider accounts.
- GLM 5.2
- An open-source large language model released by the Chinese lab Zhipu AI, described in this video as approaching Opus 4.8-level performance at significantly lower cost.
- Fable 5
- A frontier AI coding agent (likely referring to a high-capability model in the Fable product line) cited as an example of escalating frontier model costs — the presenter paid roughly $250 in API costs for 9 prompts.
- Chorus
- A cloud agent platform (chorus.com) that lets you deploy agents accessible via iMessage, Slack, WhatsApp, and Telegram, with marketplace skills and cron-style scheduling built in.
- Token budget
- A deliberate strategy for managing AI cost by routing tasks to cheaper models when frontier-model quality isn't required — analogous to right-sizing compute for each job.
Things they pointed at.
09:00toolHermes Agent
32:40toolThinking Machines (computer use)
Lines you could clip.
03:50
“He who can describe what they want the best will inherit the world.”
03:52
“The only enduring prompt hack is describing what you want.”
19:40
“If you're a good manager of people, you're gonna become a good manager of agents.”
20:10
“AI agents are kind of working their way into all of the tools that we use, almost like WiFi. It's just in the background.”
33:30
“It's the worst it will ever be. It'll only continue to get better.”
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.
analogy
00:00If you wanna get good at using ChatGPT, Claude, Claude Code, Cowork, Codex, or any other AI agent platform, you're in the right place.
00:09Today, we're talking about becoming agent native. We're the person at your company with the most knowledge about using AI agents effectively. In this video, we're gonna be covering the nine inevitable trends in the AI agent space that if you understand, will make you 10 times better at using AI agents in business no matter which agent tool you use.
00:30This video is not about the latest tips, tricks, skills, or weird prompt engineering hacks because the truth is most of that stuff will change over the next year and likely will become irrelevant. In this video, we're gonna be talking about the nine inevitable trends that won't change. Let's dive into trend number one.
00:51So the first trend that you need to understand in order to become agent native is that AI models are getting much smarter. And so when I say AI models are getting smarter, I mean that we are moving from a world of prompt hacks to just saying what you want in natural English. For those of you who were around in 2023, early twenty twenty three specifically when ChatGPT first came out, every other Twitter thread was a ChatGPT prompt hack and the most popular one was I want you to act as People would say things like, I want you to act as a virtual doctor.
01:28I will describe my symptoms and you will provide a diagnosis and treatment plan. Some people were even talking about how you should say, I want you to act as a Linux terminal.
01:38And now in 2026, nobody does this. You just ask for whatever it is that you want.
01:44I want you to act as basically does nothing. Similarly, for those of you who are around in the early cursor days, in the early days of vibe coding, you remember how important it was to at mention the file that you want to edit and people started calling this context management because you were always having to at mention the different files in your code base.
02:06And now all the best developers that I know simply just type prompts into Claude code or codecs, and oftentimes people use the voice mode and just yap into Claude code, and they just send the prompt off, and they don't spend any time at mentioning the different files because AI agents and AI models are getting so good that all it requires is natural language.
02:31The better you are at describing the thing that you want, the better AI is going to be at creating the thing that you want. You don't actually need to know the hacks. You just need to be able to describe to a computer what it is that you want.
02:43And again, this is true across the board. It's not just normal LLM work or vibe coding. It's also AI generated images.
02:50For those of you who are around, you remember that mid journey was full of hacks or these things called parameters. So if you wanted an image that was high quality, you would use dash dash quality or dash dash q. If you wanted it to be personalized, you would do dash dash p.
03:08If you wanted to use a style reference, you would use dash dash s ref and then you would type out this code that maps to a specific style. Fast forward to today, and if you're using the GPT image two model, none of that matters.
03:22There are no hacks. It's simply about descriptions. So I can click on this image right here.
03:28Please make his shirt green. Make his hair orange. I'm talking about the guy on the left.
03:32The guy on the right, give make his glasses blue. Make his hair green. Make the background logo yellow, make the entire background rainbow colored, and make the text say English is all that matters.
03:45And then I can just, you know, run this as many times as I want, and English is all that matters. There's literally no prompt hacks.
03:55Describe what you want, and there you go. All of these were generated. And I you can pause the video, look at my prompt, which was just me speaking into the g p t image two prompt, and it came out nearly perfectly for all of the images.
04:10So the lesson of trend number one is he who can describe what they want the best will inherit the world. Said another way, the only enduring prompt hack is describing what you want. And so the second AI agent trend that you need to understand is that skills are taking over and they will self assemble.
04:30And so if you use Claude Code, Claude Cowork, or if you're like me and you use Codex, you probably have set up a ton of skills. Like, I have this Excalidraw diagram skill. I also use plugins like the documents plugin or the Gmail plugin.
04:45I also have a YouTube thumbnail plugin, a YouTube researcher plugin, or a YouTube channel brief skill.
04:53All of these are skills, and all of these skills point to instructions. These are just instruction files that are stored somewhere inside Codex and these skills, you can find them by clicking plugins and you can click skills and these are all of the skills that I've created.
05:10I've created over a 100 different skills that I use every single day. And so whether you use Claude, and that includes the Claude desktop app or Claude co work or Claude code or if you use codex or cursor, all of these tools have skills and all skills are are task specific instructions for the agent.
05:31And so this is how many people started creating skills if they were using a tool like Claude and this is the way that I actually don't recommend or I actually just personally don't like doing. You can just go to the customize tab. On Codecs, it's the plugins tab, but you can go to skills and you can hit plus and you can, like, you can create a skill by uploading a skill and you can upload these skill files and so you can manually type these out and you can create these skills that your agent can use when your agent deems it necessary to use these specific instructions.
06:06And this is not the way I create skills. The way that I create skills is I will tell the agent to do a thing, then I'll get the agent to do the thing better or to make the thing better, and then I'll just tell it to turn it into a skill. For instance, I have a skill called YouTube researcher skill.
06:25Please look up my latest codex video, which is called something like learn codex in 95% of codex in twenty something minutes. I want you to create write out the exact hook from that video and then outline the rest of the video.
06:40So here's the agent's response. It extracted the transcript of the video, turned it into the full intro, and then it outlined the rest of the video.
06:49This is really good. I have no notes for this. Please, I want you to turn this into a skill called hook outline.
06:59Okay. So now it's working in the background, and it is gonna create a skill that will show up here called hook outline. So now that that's done, I'm gonna make a video on Devon, the AI tool very soon.
07:10So I can say, I'm making video on Devon. Please try to decide what I would say, and I want you to create a, uh, slash hook outline. We've created a skill and it should create it in that exact format that I just showed you.
07:24And there you go. I know this is a very simple skill, but look at this. It created it in the exact format that we specified earlier.
07:31And let's say I actually want these links as hyperlinks in the outline.
07:38I'm gonna just say, hey, please remake this. If the sources are relevant in the outline, please include it there. Then once you're done, I want you to update the skills so you do that in the future.
07:49Okay. Here. It says, done.
07:51I also updated the hook outline skill. And so in the future hook outlines, I'll put the source links directly in the relevant intro or outline beats. And there you go.
08:01We have the intro and the outline with relevant links within the actual description, not just listed at the bottom like it is here. So the takeaway for trend number two is just to use skills often especially for repeatable tasks.
08:15And the second part is when you notice where a skill can be improved, just ask. And the meta trend or the longer term trend that I want you to notice is you're gonna see this process happen automatically.
08:28Your AI agent is just going to create skills for you and then, uh, depending on your responses, your skills were will auto update. The the AI tool that does this the most is actually the Hermes agent, which is kind of just a better version of OpenClaw.
08:44Their AI agent is really good at auto updating skills where you actually don't even need to ask. And so the future of AI agents is just auto updating skills depending on how you interact with it.
08:57So trend number one, smarter AI models. Trend number two is skills are becoming really important and soon they will self assemble, and trend number three actually encapsulates both of these two trends.
09:10Trend number three is that AI agent platforms are becoming general purpose platforms. And so when I say general purpose platforms, I mean super apps.
09:20And the two main super apps that are available on the market are Claude desktop and Codex. And Claude desktop, if you can think of a super app, the three main functions are just chat, co work, and Claude code.
09:34You can chat with Codex. You can create presentations on Codex, which is just like co work, and you can also do coding tasks on Codex. And so the main takeaway is that these model providers like OpenAI and Anthropic are building platforms.
09:50So this is Codex, one of the super apps, and it has basically everything that I might need in order to do knowledge work. It can create a new chat, so I can chat with an AI agent. That's what I'm doing here.
10:02I also have all my plugins and skills that, again, I can create just by talking to my agent. I have all of the automations. So anything that I type here, can turn to an automation.
10:12We'll get to this in just a second, but you can create automations. And as of two weeks ago, they also added sites. So any website that you create with the plugin sites on Codex will actually host the app for you.
10:25It'll put it on the internet. This is a direct competitor to Replit, Lovable, Bolt, tools like that that do all of these hostings.
10:34You can just do that now inside Codex. Right now, this is only for internal tools and internal sites for you and your team, but very soon, I have a feeling that this will be for everyone. You'll be able to host your ChatGPT sites and share them with anyone.
10:48A good super app, you should be able to multitask. So at any point, can hit command n. Please do research on my email.
10:57Tell me what I need to respond to. And if I want to, I can type Gmail.
11:04We can use the Gmail plugin. And then as soon as that starts going, I'll see it here on the side panel. I can hit command n.
11:11Please tell me my schedule for tomorrow.
11:16And then as soon as that sends off, we can see that it's working here on the side. Please tell me the improvements that I need to make for my personal website. Let me know.
11:26And as you can see here, you can multitask. You can talk to the agent that's running in your super app, and you can talk to many agents in parallel. And then the most important part in my opinion or at least the part that I I'm most excited about is the in app browser.
11:40Your AI agent, because it can search the web, it can find any link. Right? Here we can see that this gave me a link.
11:47I can right click on it and hit open it in browser and it will open it up here because this is a full browser. And this built in browser is getting better every single week basically. And so that's all the general agent stuff, but they also have built in all of the vibe coding stuff.
12:02Like you can press command j and you can open up a terminal and you can run Claude. Right? You can run Claude directly inside Codex, which I do quite often because Claude is better at design than the OpenAI models, but I like the Codex interface better.
12:19I like the super app better. And so the main takeaway from trend number three that involves general purpose platforms or super apps is you should just simply learn one of the super app platforms. They are here to stay, they're incredibly fun to use, and they're incredibly useful.
12:34So the fourth trend that you really need to understand in order to become AI agent native is that these AI agent tools or these AI agents are connecting to the tools that we already use. And so on these super apps, you can very easily tell your AI agent to summarize Slack, summarize your email, or also grab everything from linear.
12:56But this is only just the beginning. Not only can you have your AI go check these data sources, you can communicate with AI agents directly through these platforms. Now you guys should probably recognize this app right here.
13:10This is iMessage, a tool that billions of people use. Well, guess what?
13:15I have ClaudeCode and Codecs directly inside Chorus, and I could say something like, please, can you analyze my latest 20 videos on my YouTube and make a personal landing page for them that I could send to potential sponsors.
13:38And so this is actually a group chat with Anj and Mohammad, and the AI agent is going to respond to me directly inside this chat. And look at this.
13:48It just responded. So this is my content agent that I created on chorus.com. Love this.
13:53Pulling your latest 20 now and building a sponsor ready page. Give me a few minutes, and here we go. We can see this page right here.
14:01It opened it up and I created this directly through iMessage. And so the general purpose platforms are super apps or AI agent platforms meant to be used on the computer.
14:13Whereas these AI agents that connect to the tools that we use where we can communicate with them through iMessage and Slack, for these you should be using cloud based AI agents. And so these are AI agents that never turn off and these are a lot better for agents you want to use in a group.
14:33Maybe you have a small team, maybe you and one other person are creating a company. It's great to use cloud based agents. So we just created an iMessage agent.
14:43Now let me show you how you can create a Slack based agent. So this is the agent, Riley's marketing agent that I created on chorus.com, and I can very easily give it a phone number to text directly through iMessage and it's literally one click.
14:58It'll give you a number. You can text that. But also, I know a lot of people for work like to use Slack.
15:03So you can go to your platforms and we've already added the SMS and here you can also add Slack.
15:11And so now I am in my Slack channel, and I can just type at Chorus. Hey. I need you to do in-depth research on Alex Formosie and how he's done such a good job with his YouTube channel, and I want you to tell me 10 ways that I can apply all of his insights or the way that he makes videos into making my content better.
15:32Also, look at Theo from t three chat, that guy on YouTube. And then I want you to make a little landing page which just kind of describes how I can apply it to my own work. When And you're done, I want you to email it to Anj and Emily.
15:46And you could see here that Chorus just reacted to it. So my agent has literally reacted to it, and so now I know that it's working on it. And there you go.
15:55A couple minutes later, we do see this response from the AI agent. And I just gave it the email addresses for those people. And now it's the Chorus agent is thinking and it should give me a link to that page.
16:08And there you go. The AI agent directly inside Slack responded with this public link, and we can open this up, and here we go. We have the Hormozi and Theo Playbook.
16:17It analyzed all of their YouTube videos and gave me this report that I can just send to people. This is public on the Internet. And just like the super apps, the same exact way you can also add skills.
16:29You can browse a marketplace of skills just like you can in the cloud desktop app and the Codex application. You have platforms where you can message it directly through iMessage, Slack, WhatsApp, and Telegram. And that's kind of the difference between these Cloud Agents and these Super Apps is because this is running in a computer in the Cloud, you can message it through all these different platforms and your computer doesn't need to be open.
16:53And again, you have all of these different connections, so you can connect it to all of your different tools and you can do cron jobs, which we'll get to in a little bit. And so trend number five that you really need to understand to become AI agent native is that foundational skills will rise to the top.
17:10So as AI gets smarter, which was trend number one, it becomes easier to create skills or repeatable instructions for your agents. As AI agents become more general purpose and as general agents, specifically cloud agents, enter the tools that we use like iMessage and Slack, foundational skills will be what matters.
17:32So you can think of it as the at the top, you have like hacks, like strategies, like we were talking about early, prompt hacks or different APIs you can use. And then in the middle you have tools. Right?
17:43You have your for the general agent platforms, you have Codex and Claude.
17:48On the cloud side, you have tools like Poke and Chorus and others. You have the tools, but what actually matters is not the tools that you use, rather the foundations that you have as a person.
18:01And a lot of these are communication skills. Because agents are gonna start to feel like just talking to a team of AI agents in platforms that are extremely intuitive, partly because super apps are gonna be really easy to use, and then also AI agents will just live in the tools that almost all humans already know how to use, What matters is just communicating with these agents.
18:22Can you communicate with Claude? Can you communicate with Codex?
18:27How are your delegation skills? Part of that is understanding what good looks like, and so that means you need to become an industry expert at whatever it is that you're doing. I'm really good at creating content.
18:36That's kind of my superpower. I am definitely able to delegate an AI agent to do research on content or to create a thumbnail or to create anything else that has to do with content because I understand what good looks like.
18:49If you asked me to create a discounted cash flow analysis on a company, I would not be good at delegating to an AI agent because I don't know what a good DCF analysis looks like. On some of the other like soft skills, like mental clarity and multitasking, If you are already good at this thing, like managing people and doing all of these things, you're gonna be good with AI agents.
19:12If you're a good manager of people, you're gonna become a good manager of agents. And that's because communicating with AI agents is gonna feel like communicating with a human very soon, except they're going to be working twenty four seven and you're gonna be able to add skills to them. You can't really do that to a human as easily as you could do it with an AI agent.
19:30And so the point I'm trying to make in trend number five is that AI agents are kind of working its way into all of the tools that we use, almost like WiFi. It's just in the background. You have a team of AI agents always working.
19:42And in this new world where you have agents working, foundational skills will rise to the top, and I recommend getting good at these five things. Trend number six that you need to understand in order to become AI agent native, AI agents are going to work asynchronously. Meaning you can take anything that you type into codex or Claude and you can turn it into an automation.
20:07So if you remember earlier, we created this hook outline skill and then we used it on Devon. And what if I wanted an AI to do this every morning and I wanted an AI to actually come up with the best possible YouTube idea and then I want it to create a hook outline. And this is just something I can read in the morning that might give me a good idea for what content I should make.
20:29All I need to do is tell Codex to do that. Hey, I want you to do research on the best possible video that I could make every single morning at 9AM and I want you to send me one of these hook outlines. Create this automation.
20:43And one thing you need to realize with automations is you can kind of act in the future. Anything that's useful now, you should automatically think to yourself, would this be useful on a recurring basis or would this be useful at a very specific time? The same way you can tell it to do something every morning at 9AM, you could also do it tell you to do it four times.
21:02You could say, please do it next Thursday, next Friday, the following Thursday, and the following Friday. You can be very specific or you can make it recurring.
21:09And AI, because it's just like talking to a human, will just set up the automation for you. And you can see here that it created this automation daily best video hook outline. We can go up to the automations tab and there it is, daily best video hook outline and here is the description of your automation and you can manually change it, you can tell your AI to change it, and you can test it by clicking run now.
21:34And so the way I kinda look at this trend is in the old world of AI, you had tools like Zapier where it was kind of this niche skill to learn automations, learn AI automations so that you can set up a trigger and some advanced workflow. But this new world, we're moving to a world where you can just communicate with AI agents in natural language.
21:54So anything that you do inside Codex, you can say just do this on a trigger. Every single time this thing happens, just do this, and AI agents are able to do that.
22:04You can set up a schedule, you can set up a routine, and whether you're in Claude, whether you're in Codex, you can create these automations with natural language. You can also do this on Cursor.
22:15Cursor is very similar to Codex except it's more for developers. But again, there's this automations tab. In order to create an automations, simply tell your agent to set up an automation.
22:26It's that easy. All AI agent platforms are becoming automation platform. The seventh trend that you need to understand in order to become agent native, and this is probably the coolest and most futuristic one, AI agents can fully control your computer.
22:41So I could go to a tool like Codex. Codex has the best computer use tool in the world. It's built directly inside Codex, and this will allow your agent to fully control your computer.
22:51After this video, what I want you to do is I want you to just try this computer use. Tell it to do something on your machine. AI agents are really good at controlling your computer.
23:01Now sometimes they're a little bit slow, but you'd be surprised at how fast it can do certain tasks. I believe that within twelve months to eighteen months, AI agents will be better than humans at controlling a computer.
23:15And computer use is really useful when one of the plug ins don't fully work. Right? You can use something like Gmail, which uses their servers behind the scenes and it allows you to access the functionality of Gmail without controlling the interface that is your computer.
23:33But not every task that you might wanna do on a computer has a perfect plugin. Maybe sometimes your AI agent that you're using in Codex might actually be able to just open up a browser tab or open up a file on your computer or look in your downloads. It's sometimes really useful to allow your AI agent to just fully control your computer.
23:53It can click, it can type, it can do all the things that you might wanna do on a computer. Another way that it's sort of controlling your computer is it also has browser use. And so it can either use your Chrome browser use, it can control Chrome, or it can control the browser that's inside Codex, meaning it'll just open up the browser.
24:14And remember, this browser right here inside Codex is becoming just like Chrome. It is a full browser built into Codex. The ways in which it can use your computer are computer use, in app browser use, and external browser use like Chrome.
24:29So it can open Chrome and control it like it's a human. Okay. So trend number eight that you need to understand to become agent native comes in two parts.
24:38The first part is that Frontier AI agents are getting way more expensive and it's time to token budget. With the latest release of Fable five, companies are realizing how expensive AI models can truly be.
24:53For instance, I used Fable to build a mobile app. It was actually like a lovable clone that I built using Fable five, and this was nine to 10 prompts. If you were to have done this billing through the API, which is how much this model will cost normally, it would have cost it around $250 to generate all the code.
25:14Yes. This generated for a very, very long time, but it was still only nine prompts, and it's not like it went overnight. And so just for those nine prompts, it was around $250.
25:25And the second part of this trend, while the frontier models are getting really expensive, open source models are getting significantly better, and the best open source models are coming from China. And then yesterday, the best open source model in the world was released. It's called GLM 5.2, and it is incredible at front end design.
25:47And many people are saying GLM 5.2 is just slightly worse than OPUS 4.8. And this is not just in design.
25:56This is across the board. GLM 5.2, which is significantly cheaper than Opus 4.8 is almost as good.
26:05And I'm gonna leave you with some practical things to do. So I'm just gonna show you how you can use GLM 5.2 directly inside Cursor.
26:14And so over the last forty eight hours, I've just been testing GLM a ton. And you could see here in my model selector, I have this option z dash a I slash g l m 5.2.
26:26Hello. What model are you? I am able to use cursor and use g l m 5.2.
26:33And you can see here it says I'm g l m 5.2. And so I'm gonna show you exactly how you can set this up. Okay.
26:39So by the way, this is not sponsored at all by OpenRouter. This is something that you should definitely understand if you wanna become fully agent native, and that is how to use APIs, specifically OpenRouter because it allows you to use all of the different APIs with a single API key.
26:55How can you use all of the different models effectively? If you go to OpenRouter and you create an account, you can get an API key to use any model.
27:04And once you get your API key, you can go to Cursor, and the steps are relatively easy. In Cursor, if you go to settings and you click on models and you go down to API keys.
27:17You're gonna paste your API key right here, and you're gonna enable this. And then you're gonna enable the override OpenAI base URL, and you're gonna paste this right here exactly as is.
27:29Once you do that, you need to go to view all models and then you're gonna add a custom model and then you're gonna paste this exact string z dash a I slash g l m dash 5.2 and press add. Once you do that, you will be able to use z dash a I slash g l m 5.2, and this is the best open source model in the entire world.
27:51And I guarantee you, you are going to be surprised at how powerful it is. And so this video is not about g l m 5.2. This video is about the trends, and the trends are very clear.
28:01The frontier models like Fable and whatever models OpenAI releases in the future, these do not appear to be getting cheaper. However, what does seem to be happening, these Chinese open source models seem to not be getting more expensive, yet they're catching up with the frontier AI models.
28:18And so if there's one takeaway from this trend, it would be to use a tool like OpenRouter so that you get access to all of the different models and test the open source models and see how they compare to all the workflows where you're using g b t 5.5 high and Claude Fable. There may be a ton of tasks that you do where you can just use an open source model and save five times the amount of money.
28:40And the ninth and final trend is something that I have to just show you, and that is real time voice paired with it controlling my computer. Before this video, in about three prompts that did take around twenty minutes to generate with Codex 5.5 on Codex, I created Riley's Jarvis.
28:58This is a little desktop app that I can open up right here. Let me show you something. So I'm gonna just go ahead and I can turn this recording on.
29:07Hey, Jarvis. What I need you to do right now, can you open up the Comet browser?
29:15Done. Can you open up a new tab? I want you to go to LinkedIn.
29:22Got it. Let me open a new tab and head to LinkedIn. I opened a new tab and loaded LinkedIn.
29:29Very cool. Can you actually I don't need LinkedIn. Can you open up Cursor for me?
29:36Opened Cursor.
29:38Very cool. Can you type a prompt down there at the bottom where it says send follow-up? I want you to just ask Cursor to make a very simple landing page.
29:46Can you do that for me?
29:49Alright. I'll drop a quick prompt into that follow-up field now.
29:54Search the prompt in. Thank you for typing it. Can you enter that in for me?
29:59Sure. I'll send that now. Enter.
30:01Nice. Thank you. Can you actually go to Notion?
30:06Switching over to Notion. Open to Notion. Thank you.
30:10Okay. Yeah. So there I can see the yeah.
30:12What do you think of my topics for the week, as you can see there? Let me take a closer look at what's on the screen for that.
30:19I can see topics like Mythos timing, Codex update rumors, GLM 5.2 cloning Codex, and and Cursor's acquisition. Nice.
30:27Nice. Nice. All I'm trying to show you right here is that AI is getting significantly better at allowing you to speak to it in real time while using tools, and you're able to just fully control your computer while speaking.
30:40And another company that's working on this is called Thinking Machines. And they did another demo and it actually looks better than GPT real time too, but this is one of the most fun trends.
30:53You can just speak to your computer and it can just take actions for you. And and it's the worst it will ever be. It'll only continue to get better and it'll get significantly faster and very soon it's gonna feel like Jarvis.
31:05You're gonna be able to control your computer and this is only just beginning. Alright guys, there you have it. We covered a lot today.
31:12And just a quick recap, uh, for trend number one, we talked about how AI models are just getting smarter, prompt hacks are going away and saying what you want is in. Skills are becoming significantly more important and you can create skills with natural language. Just tell your agent to create a skill and soon over time AI agents will self assemble these skills.
31:31Number three, probably the most important trend so far is that we're seeing these general purpose AI platforms or agent platforms that I call super apps, and you should absolutely learn one of these platforms. Whether it's Codex, Clawd Desktop, or Cursor, these platforms are incredibly important to learn.
31:48Number four, agents are connecting to the tools we use. Right? Not only can you use these agent platforms on the computer, you can also start to use these different cloud agents directly in the tools you already use whether that's iMessage or whether that's in Slack.
32:03You can very easily set up one of these agents at chorus.com. Number five, foundational skills rise to the top. I noticed the people who are best with AI agents are just the ones who are really good at like these soft skills or these foundational skills like communication, delegation, understanding what's good looks like or taste.
32:21In other words, mental clarity, multi tasking. It's very important. Obviously, AI agents are becoming more asynchronous.
32:27All you have to do is ask your agent to do things on a schedule and boom, automation. Also, AI agents can control your computer. And I showed on Codex that is the best platform for using computer use.
32:38I highly recommend trying it out. It can fully control your computer. It can even do it in the background.
32:42Trend number eight is that Frontier AI agents are getting more expensive. This show is showing no sign of slowing down. Enterprises are willing to pay basically unlimited money for the best AI models in the world.
32:52And, unfortunately, with the amount of compute we have, it's just not feasible to get these for cheap. They are going to be expensive for a while.
33:00But luckily, the opposite end of that trend is that local models like GLM 5.2 are getting significantly better for way less money, and I highly recommend you start looking into that. And then finally, I showed you Jarvis. Real time voice is getting better.
33:14Real time voice compared with computer use is really fun. We're in the early stages of this, but pretty soon, we're gonna be controlling our phone and our computer with our voice. AI agents are gonna be able to take action in the background, and they're also going to be able to take action directly on our computer.
33:29It's gonna be like Jarvis with like a superhuman AI assistant in the background. It's gonna be awesome. Thank you guys so much for watching.
33:35I hope you got a ton out of this, and I'll see you here for the next
The Hook
The bait, then the rug-pull.
The title promises nine skills but delivers something more durable: a structural map of where AI agent platforms are heading, and why the builders who already know how to communicate, delegate, and recognize quality output will compound fastest — regardless of which tool wins.
Named ideas worth stealing.
18:40model
The Foundations Pyramid
- Foundations (base): communication, delegation, domain expertise, mental clarity, multitasking
- Tools (middle): Codex, Claude Desktop, Chorus, Poke
- Hacks (top, perishable): prompt tricks, specific APIs
A three-tier hierarchy for thinking about what to invest in when building AI-agent fluency. Foundations are durable; tools are medium-term; hacks are disposable.
Steal forFraming any 'how to get good at AI' content — the pyramid is a clean visual that holds up across any tool shift
13:50model
Super-App vs. Cloud Agent
- Super-apps (Codex, Claude Desktop): run on your computer, best for active work sessions, built-in browser + terminal
- Cloud agents (Chorus, Poke): run 24/7 in the cloud, best for async tasks, live in Slack/iMessage/WhatsApp
The two categories serve different jobs — choose based on whether you need the agent present (super-app) or always-on in the background (cloud agent).
Steal forHelping a team decide which agent setup to invest in first
00:49list
The 9 Inevitable AI Agent Trends
- 1. Models get smarter — plain English beats prompt hacks
- 2. Skills take over — reusable instructions, self-assembled over time
- 3. Agent platforms become super-apps — Codex, Claude Desktop
- 4. Agents connect to existing tools — Slack, iMessage, WhatsApp
- 5. Foundational skills rise — communication, delegation, taste
- 6. Agents work asynchronously — natural-language scheduling/automations
- 7. Agents control your computer — click, type, browse autonomously
- 8. Frontier costs rise; open-source catches up — token budgeting matters
- 9. Real-time voice + computer control — the Jarvis moment
The nine structural shifts the presenter believes are non-negotiable regardless of which tools win.
Steal forListicle content framing, trend talk structure, onboarding new team members to AI agent thinking
How they asked for the click.
VERBAL ASK
32:45next-video
“Thank you guys so much for watching. I hope you got a ton out of this, and I'll see you here for the next”
Soft verbal sign-off with no hard subscribe ask and no product pitch — consistent with educational positioning throughout.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
