That's what i discuss today with @rauchg.

As well as:

- How to get started building agents?

- Do we build a team of agents or one GOD agent?

- OpenClaw, and the movement it started.

- Hard parts of creating agents

- Agents that delegate to other agents with different permissions and computer access

This is @agentnative_ #4

00:00 Intro

01:21 The State of Agent Adoption

07:26 "V" Vercel's Internal Company Agent

10:18 Building V and the Eve Framework

14:32 Lessons from OpenClaw: Soul MD + CPU use

22:20 One God Agent OR a Team of Agents

27:11 How to Build Your Own Agent: Skills, Tools & Permissions

33:52 Proactive Agents, Triggers & Self-Improvement

40:30 Who Will Build Agents & Getting Started

46:22 The Future: Cheaper, Faster Models & Open Weights

53:05 Closing Thoughts: Owning Your Own Intelligence

Here’s a weekly agent native update for anyone looking to implement frontier AI agent tools and models into their business.

Long story short: Some Codex and GPT Work updates, DeepSeek is back, frontier intelligence is getting cheaper, tools like Buzz are becoming collaborative workspaces for agents, and Grok and Meta are building vibe-coding tools that will integrate directly into their social platforms.

Buzz puts humans and agents on the same team:

Buzz is beginning to gain attention - for good reason.

• Buzz is a new way to interact with AI agents. Buzz is a free Open Source project created by Block + @jack.

• It is the first experience i’ve had working with AI agents in what feels like a group chat.

• Every agent has its own identity and permissions, allowing people to interact with agents more like teammates than background bots (Like Slacks current agent configurations)

• You can sign into buzz with your existing subscriptions

• A complete guide to getting started can be found here:

Buzz has a desktop app that allows you to interact with Codex and Claude Code in what feels like a group-chat.

But they also have a mobile app that is VERY fun to use. It acts like the Codex remote in that you control the desktop app remotely.

This is what the iOS app looks like.

The entire Buzz project is available on GitHub.

OpenAI made GPT-5.6 significantly cheaper

OpenAI cut the price of GPT-5.6 Luna by 80% and Terra by 20%

• If you use Codex, the best way to conserve tokens is to use Terra and Luna as much as you can.

• You don’t need Sol for most things, btw.

• My thoughts - Terra is a solid model for everyday professional work (Good for most GPT Work Tasks).

• GPT-5.6 Sol also received a Fast mode that runs up to 2.5 times faster at twice the API price, without changing the model’s intelligence. You will run through your limits pretty fast using this - this is the double edged sword. The faster the models get, the faster the token burn.

• OpenAI explains the full pricing update here

Codex has some new updates

Recent Codex updates make it easier to use agents as part of everyday work.

The first update is the new notifications side panel. Its much easier to use and stay organized than the previous view which didn’t show the folder you we’re working in. This feels much more natural.

The new Codex Chrome Extension:

The Chrome integration also allows Codex to work across websites and tabs where the user is already signed in. I usually use the Codex Browser for most things, however its nice to have inside chrome. These chats are synced with the Codex app so your chats will show up inside Codex, when you use the Chrome Extension.

Grok added Vibe Coding... (Meta is as well)

Yep, Grok is adding Vibe Coding to the platform:

This is an app that lets people:

• describe an application

• have Grok build it

• and publish the result on its own domain

Meanwhile, Meta launched Pocket, a feed where people create, play, discover, and share small interactive apps and games generated from prompts.

The reason I’m including this is because I think it’s relevant to businesses. The nature of SaaS is changing a lot. Software is starting to become more like content, something that may only be relevant for a few days. Games, mini productivity apps, and maybe soon... much bigger mini apps, integrated directly into social platforms.

There’s been a lot of conversation lately about whether SaaS is dead. I don’t think it’s dying, necessarily... I think the way we consume software is could be changing.

When I was a kid, I bought individual songs on iTunes. Then Spotify came along and gave us access to nearly every song in one place.

Maybe software is heading toward the same shift. Instead of buying dozens of individual tools, we’ll access a platform that can create whatever software we need, whenever we need it.

So what will become the Spotify for software?

Maybe it’s Grok Build.

DeepSeek dramatically changes the API price comparison

DeepSeek just released its updated V4-Flash API. It’s been improved specifically for agent work, supports the Responses API, and is adapted to work with Codex.

The interesting part is the price. Compared with Claude Fable 5, DeepSeek is around 71 times cheaper for input and 179 times cheaper for output.

The model is showing sings of being close to as good as Opus 4.8, which is insane or the price.

Riley’s Thoughts

Pretty slow week all things considered. No huge new models, no huge updates to claude, codex, cursor or any other huge agent platform.

But the moral of the story, whether you’re using American models or Chinese models... Intelligence is getting cheaper. Which is always fun. I will do more testing on the Deepseek models and will report back next week.

Subscribe to Vinny https://youtube.com/@stealtime?si=khW3yG-ovHdjYAtY

Buzz is a new free platform, that feels exactly like Slack,

except the agents aren't add ons, they are equals.

and in 5 minutes you can create an agent team.

And it works with your existing Codex and Claude Code subscriptions... So it's free to use.

This video is divided into 2 parts:

---

⭐ Part 1 Conversation with Vinny (@hot_town on X)

00:00 Introduction

01:59 Why is Buzz Going Viral?

07:08 Creating Agents & Channels with Fizz

20:03 Adding Other Harnesses (OpenCode, Cursor)

21:25 The Buzz Mobile App

22:36 Payments & Shared Compute

25:36 Workflows & Projects: A GitHub Killer?

30:14 The Future of Buzz: An Open Cowork?

35:13 Automations, Triggers & Trusting Your Agents

---

⭐ Part 2 - Getting Started Tutorial (My workflow)

42:32 Tutorial and Breakdown

43:03 Downloading Buzz And Default Agents

43:35 Creating your first agent

45:13 Adding Agents to channels

45:43 Why I use Codex as Lead Agent

47:06 Using Codex, Claude Code and Grok for task

48:52 Adding Muse (Meta Model) using OpenRouter

51:23 iOS app is Free and easy to use

51:50 My management Channel (My #1 use case)

In this episode, I sit down with Ras Mic to break down what this means, why every company should be building agent-native software, and the Codex workflows we’re using right now. We cover computer use, agent loops, automations, multiple threads, the tools that make agents more powerful, and how Codex compares to Claude Code and Fable 5. T

his is the beginning of an entirely new way of using computers—and most people still have no idea what’s possible.

Check out Ras Mic’s channel: https://www.youtube.com/@rasmic

00:00: Introduction

00:40 OpenAI Merging ChatGPT and Codex

02:20 Sol, Terra, Luna - The new Models

05:05 The Replit/Lovable Benchmark (Fable and 5.6)

08:20 Pricing of 5.6 (it's great)

09:00 Computer Use

15:10 Record and Replay for knowledge work

16:15 Loops

20:20 Loops for Marketing

26:10 Plugins and Social Media Scraping

28:20 Tools

30:45 Building Block Economy (Legos for Agents)

34:05 Use Multiple Threads

37:20 New Sites Feature 40:00 Updated Browsers

41:50 Agent First Operating System

• Andrej Karpathy joins Anthropic

• Claude Agent View in Terminal

• Claude Desktop improvements

Claude added a new Agent View in the terminal, which lets you spin up multiple Claude Code agents at once and manage them from one screen within your terminal

Karpathy joining Anthropic is a massive deal, not just because he is famous, but because it shows how hard these labs are fighting for the best people in AI.

Claude’s biggest issue is product clarity, because Claude, Claude Code, Cowork, skills, and the desktop app all feel a little too separated right now. Example: Skills you create in cowork don’t work in the Claude Code app.

OpenAI / Codex

• /goal now available in the Codex app

• Share Plugins with Teams

• Design Mode / advanced annotations

• Appshots

Codex added existing feature /goal to desktop app (It shows up in the UI), which basically lets you give Codex a long-running objective instead of just a normal prompt. Some people have reported the agent can work for over a day.

Plugin sharing just released. Big deal because teams can now package up skills, integrations, and repeatable workflows, then share them across the workspace. This will be a big deal over the next 6 months as teams migrate to super-apps.

My favorite update is Appshots, because you can press both Command keys, send whatever app you are looking at into Codex, and immediately give the agent real context.

Google / Gemini

• Gemini Spark

• Google I/O 2026 updates

• Gemini 3.5 Flash / product stack updates

Google I/O had a lot of AI agent talk, but the one thing I keep thinking about is Gemini Spark, which is their new always-on agent inside Gemini that will soon be on the gemini app.

Their new model was mid at best.

The problem with Google is I still do not know what Google’s actual super app is... Is it Gemini, Antigravity, and AI Studio? Why separate the agents when the magic of agents are that they eliminate the imaginary boundaries between types of tasks.

GOOGLE.... MAKE A SUPER-APP. COMMIT TO ONE.

Cursor

• Composer 2.5

• Cursor Automations

• Cursor mobile app on the way

• Cursor Automations changelog

Cursor might be back in the race, because Composer 2.5 is really fast and feels especially good for frontend and vibe coding work. Many are saying it’s close to Claude Sonnet 4.7 in terms of coding ability.

Cursor announced Automations into the Agents Window, which makes Cursor feel less like a code editor and more like a place where you manage agents.

With mobile on the way and the SpaceX partnership giving them access to way more compute, Cursor is clearly trying to become a full AI super app for coding and knowledge work... Like Claude Desktop and Codex

See you next week

The ironic thing about the AI bubble bursting is that it will actually cause more people to use AI.

If NVIDIA is truly overvalued, and it drops 60% and the entire circular AI economy crashes before our eyes, this will cause a massive round of layoffs in big tech and other industries.

And this will actually cause these companies to want to be more efficient with their capital.

And since AI is objectively useful in coding, image generation, image editing, simulations, robotics, AI generated music, video generation, video editing and hundreds of other use cases, companies with smaller budgets will be forced to use AI tools instead of hiring humans.

And so my advice to you in these bubblish times are as follows:

Go play in the bubble. This one is fun.

You know, 4 Months ago I raised 10 million dollars to build the capcut for app development, and every single month that goes by, a new model comes out that makes our app much better and much easier to create apps directly from your phone without coding experience.

The rate of progress of AI in terms of objective utility is INSANE.

Look up the Google Gemini 3 and Nano Banana Releases of the past few days.

AI is literally developing a visual understanding of images, designs, concepts and simulations...

And AI is not like crypto or NFT’s. Because Crypto and NFT’s we’re both cultural phenomenons.

They both become worthless if no one agrees they have value.

But AI is objectively useful right now in isolation, even if all progress on AI stops, and you’re the only that uses it.

Whether or not the macro valuations on AI are too high... the micro implementations of AI are only going to RAPIDLY increase, whether or not a “bubble bursts”

When people who hate AI talk about the bubble bursting it seems to come from a place of fear and bitterness surrounding AI.

They almost want the bubble to burst. As if it’s going to help them.

And I hate to be the one to tell you this... but that’s just not the case.

If anything... companies are avoiding going all in on AI because the models keep getting better, which makes it hard to decide what to focus on.

A bubble burst would indeed slow down the progress of frontier models which would force companies with now smaller budgets to focus their efforts on practical ways to use AI.

So my advice to you. Focus your efforts on practical ways to use AI.

Bubble or not, you’ll be in a great position.

Use the Vibecode App: https://vibecode.go.link/lHp66

In 14 minutes: - I Built a Mobile app From my Phone using Claude 4.5 Sonnet on vibecode.dev

- I designed and modified my app

- I sent the app to the Apple app store

TIMESTAMPS:

00:00 Introduction

00:21 Defining the Problem and App Idea

01:39 Building the Initial App

03:24 Fixing App Issues

04:50 Improving the Design

10:21 Final App Features

12:29 App Store Submission

-test out Anthropic's new Claude 4.5 Sonnet model and its new interface

-we compare Claude and Codex,

-test out the new model by creating a browser based face filter app using Snapchat's face filters as an example

-we also built a fun donut visualizer inspired by Lex Friedmans donut video from 5 years ago

If you're interested in trying out this model uninterrupted and without needing to set up a technical environment, I can't recommend trying out https://sandbox.dev/home enough

and if you haven't had a chance to try it out yet, and have ideas for an iPhone app but no technical knowledge, i'd highly recommend trying out my app Vibecode: https://vibecode.go.link/lIlvc

AND to get you started with all of this, I'll send you a free guide through the link below.

This guide teaches you how to use both Claude Code and Codex to build different types of apps: https://maven.com/p/9c611e/claude-code-and-codex-guide

TIME STAMPS:

00:00 Introduction

00:34 Setting Up Using Claude 4.5

01:36 The App Idea

02:40 Building the Face Filter App

05:41 Context Anxiety

08:00 Why Using Sandbox is Better

13:24 Parallel Processing Capabilities

16:24 Creating Custom Masks with the Filter App

18:57 Lex Friedman's Donut's with Claude 4.5

26:46 Exploring Changes to Claude Code Interface

27:30 Final Thoughts

Where to find us:

Muhammad on Twitter/X https://x.com/zzulali

https://x.com/rileybrown_ai

vibecodeapp.com

Use Claude Code and Codex in the web: https://www.sandbox.dev/

In this video I go head to head with Vishal (https://x.com/vishal_dubey1)

Who can build a better video game in 7 prompts?

Vishal is using Claude Code I am using Codex with GPT-5

We're using these tools in the @sandboxvibe so we can just use them in the web and immediately share the links, and play each others game.

To vote for winner RT and comment below who you think won. (Prize: Selecting 5 people who vote to win $1000 worth of @sandboxvibe credits)

TIMESTAMPS:

00:00 Intro

01:01 Vishals Stack (Claude Code)

02:51 First Prompt with Claude Code Star Wars Game

04:19 What is "UltraThink" on Claude Code?

05:39 Rileys App Idea for Codex with GPT-5 High

07:33 Why Codex is the best Model in the world for Coding

08:16 First Prompt with Codex on Sandbox (Riley's Stack)

11:13 Vishal Testing his app

12:16 Prompt 2 - Simplifying Claudes Over Creativity

15:17 The Game Codex Created

15:57 Riley Prompt 2 - Making Changes to Characters and Cars

17:07 Creating Github Repository for Game Assets and Later Use

20:07 Adding Ability to Shoot Lasers

22:54 Prompt 3 Riley Testing Sound Effects

23:39 Adding More sound effects and changing the characters in game

28:41 Vishal Added Enemies - Testing

30:21 Prompt 4 - Adding Health Bars and making movement more realistic (Vishal)

33:34 Adding Real Time DB and Auth

37:28 Refining Controls

40:39 Riley Testing Database and Auth with Codex

41:20 The Multi Player Game Works!

43:36 Fixing MultiPlayer and Game Creation issues and adding Grenade

45:54 Advice on Database

46:32 Vishal... 2 Prompts left

48:13 Fixing Aiming Issues

49:41 Vishal Compacting Claude Code

52:09 Fixing the Grenades Spawning Location

52:48 FINAL PROMPTS

58:27 Adding Health

59:15 The Games are done: Final Testing

01:04:47 Final Thoughts and Chance to Win $1000 Worth of Credits

Get Access to Claude Code, Codex and the Mobile App Builder: https://www.sandbox.dev

Here is a complete guide. L

et's build 13 Apps using the best coding tools in 96 minutes.

TIMESTAMPS

00:00 Intro

04:15 What are these tools?

05:30 Learning Claude Code

06:06 First Prompt in Claude Code

07:02 Editing Prompt with Claude Code (Making a linktree)

07:52 Prompting Codex w GPT-5 and Gemini CLI

09:27 Running on Skip Permissions Mode (Yolo Mode) by default

10:42 Commands in Claude Code /cost /model /clear / compact

14:24 Resume Previous Chats with /Resume in Claude Code

15:32 Sharing web app links immediately with sharable links

16:20 Building Static Pages (Vibe Marketing)

18:45 Vibe Marketing 1: landing page in Claude Code and Codex

20:28 Evaluating the Landing Pages with Claude Code, Codex, and Gemini

21:00 Making edits to our landing pages + Context Management + Wispr Flow

25:47 Vibe Marketing 2: Creating a Lead Magnet with Claude Code, Codex w GPT-5

26:56 Using Tally to embed into our Lead Magnets

33:30 ‌ 34:03 Vibe Marketing 3: Creating a research Presentation with Codex

37:05 Polishing and editing our presentation

38:25 Vibe Marketing 4: Building Prototype, Analytics Dashboard, and Game at the same time with Claude Code and Codex

39:20 Using Canvas app to mockup for prototype (Excalidraw)

42:23 ‌ 47:38 Dealing with an error in Codex w/ GPT-5 high

50:34 Vibe Marketing Recap... What we created

51:46 PART 3: Building Notes App with Database and Authentication using Instant DB

52:13 Using InstantDB to build apps

01:02:21 API's (AI Power ups for your vibe coded apps)

01:02:35 Building Thumbnail Editor App with Nano Banana API

01:07:54 Bringing a Prototype to life with Nano-Banana

01:10:09 What are API Keys? Why do you need them?

01:12:41 Getting Google API key for Nano Banana

01:19:33 Building a Mobile app

01:21:42 Building a mobile app using AI API's

01:30:17 Deploying Mobile apps to the App Store

01:33:53 Deploying your apps from Sandbox

IN THIS VIDEO:

1. I'm giving a breakdown of what Claude Code's new Subagents feature is

2. Creating a master agent that delegates tasks to specialized subagents

3. Demo how these agents can operate in parallel

4. See how they handle different tools, including API integrations with X and FAL for social media postings and image generation

ClaudeCode has subagents for the web but if you want to build your own version of this on your phone, you should definitely check out VibeCode: https://vibecode.go.link/1kMAR

TIME STAMPS:

Intro 0:00

00:31Sub-Agents Overview

08:52 Sub-Agents Demo (Getting Started)

And you can recreate popular apps without writing a single line of code.

And ship them to the App Store:

IN THIS VIDEO

Ansh Nanda (@anshnanda_) and I:

1. Plans and Outlines his app idea

2. Builds a mobile app on @vibecodeapp

Try out VibeCode: https://vibecode.go.link/eUk3V

3. Connects to Cursor

4. Creates personalized AI-generated images

5. Submits the app to the App Store

TIME STAMPS:

00:00 Introduction

00:30 The App We're Building

01:32 Creating the App Mockup on Excalidraw

04:23 First AI Prompt on @vibecodeapp

08:14 Enhancing App Functionality

10:13 Connecting the App to @cursor_ai

12:03 Styling and Final Touches

19:26 Publishing the App to the App Store (in less than a week you'll be able to push to the app store directly from VibeCode app)

In this video we discuss:

1. How to install and set up Claude Code on your computer.

2. Using Claude Code to create basic files, do research, and generate content.

3. Building a landing page and deploying it to the internet using GitHub and Vercel.

4. Adding advanced features like user authentication, a database, and AI chat to your web app.

5. Integrating your projects with tools like Firebase and OpenAI for seamless development.

Claude helps you build. VibeCode helps you launch. Try it → https://vibecode.go.link/cRtDc

00:00 Intro

00:44 Installing Claude Code

02:12 This is Claude Code

03:39 Creating Folders and Text Files

04:35 Using Claude Code as a Content Research Agent

16:12 We're creating a very predicatble agent

18:17 Review: It's a general agent, that can control your cpu

19:13 Building Apps with Claude Code

21:47 Bypassing Permissions (Dangerously)

22:56 Creating a landing page

25:19 Setting up and "Saving" to GitHub

30:38 Deploying to Vercel

32:44 Building a web app with auth and db (Firebase)

35:14 Setting up Firebase

39:09 Our Web App With Auth Works!

43:06 You can add your own domain

If AI agents can browse the internet for you, post on LinkedIn, and summarize your research… what’s stopping you from building your own AI-powered app? 👉https://vibecode.go.link/iZ8py

In this video, I compare the top AI-powered browsing tools: Perplexity, Comet, and DIA to show how fast agentic browsers are evolving. We walk through real examples of how these tools can click, search, write, and even post for you across the web.

We cover:

-How agentic browsers like Perplexity and Comet actually work

-What OpenAI’s rumored browser might include

-The rise of deep research + browser use agents

-Why this shift could completely change how you work

-What it means for startups, creators, and SEO in the AI era

Where you can find more of my content: Follow me on X: https://x.com/rileybrown_ai

TIME STAMPS:

00:00 Introduction to AI Powered Browsers

01:42 Perplexity Browser Features

02:19 AI Agent Capabilities

03:27 Comparing DIA and Perplexity

04:01 OpenAI's Upcoming Browser

07:18 Operator and Deep Research

08:48 Skills, Spaces, and Projects

11:02 User Growth and Market Trends

11:50 The Vibe Browsing Revolution

14:24 Future of AI Browsers

15:29 ChatGPT as a Business Tool

22:45 AI's Impact on Jobs and Expertise

26:45 Adapating to AI Powered Tools

29:26 Conclusion and Future Insights

As always, thank you very much for supporting me by listening to this podcast. If you liked it, please hit download, subscribe and check out the other podcasts on the channel. I’d love to hear your thoughts or ideas for future topics—drop them below in the comments. – Riley

DOWNLOAD VIBECODE: https://vibecode.go.link/J7zUg

Where you can find more of us:

Follow me on X: https://x.com/rileybrown_ai

Follow Muhammad on X: https://x.com/zzulali

Here is the entire process start to finish in 47 minutes.

In this video, Muhammad and I attempt to fully recreate a viral iPhone app that made $7M last month — all using AI. We built it entirely inside the VibeCode app, across iPhone and iPad, without writing a single line of code. From swipe-to-delete logic to App Store submission, we go from idea to launch in one session.

TIMESTAMPS

00:00 Intro

00:29 App Idea (We found on @SensorTower+ TikTok)

00:57 Initial AI Prompt to build the app 02:18 Building on iPad and iPhone w/ @vibecodeapp

03:33 Testing native photo access and deletion

08:03 Adding Tinder-style swipe interface

13:26 Fixing animation and UI bugs

27:37 Making the Folders Feature Better

30:54 Creating app icon and landing page with AI @ChatGPTapp

33:48 Adding splash screen while photos are loading

36:24 Putting the App on App Store / SSH - @cursor_ai41:40 Preparing for App Store submission