22:26David Ondrej · Tutorial
This MCP makes Hermes Agent 10x more powerful
A 22-minute live walkthrough of wiring Hermes Agent to Apify MCP connectors and Supabase to automate lead scraping, scoring, and outreach.
June 15thThe argument in one line.
The bottleneck in agentic work is no longer the model or the cost of tokens, it is the human who has to start and unblock every thread, so the real job becomes designing triggers and loops that run without you.
Read if. Skip if.
READ IF YOU ARE…
- A solo founder or small team lead who already runs several coding agents a day and keeps hitting the wall of being the one who has to unblock them all.
- Someone paying for a $200 a month coding subscription who finishes the week with most of the allowance unused and wants to turn that capacity into output.
- An operator deciding whether to keep agents on a local machine or move them to cloud sandboxes, who wants the security and scaling argument laid out.
- A developer who has built a few skills and prompt templates and wants the next step up, which is scheduled and event driven automation.
- Anyone trying to work out what their job looks like when the task level work they do today is handled by agents.
SKIP IF…
- You have never run a coding agent. The whole talk assumes you already use one daily and are frustrated by its limits.
- You want a step by step tutorial. The live demo is a product walkthrough of the presenter's own tool, not a vendor neutral setup guide.
- You need a stable model comparison. Specific model names and pricing claims here are tied to one week in October 2026 and will age fast.
- You are looking for proof rather than argument. Most of the forecasting is reasoned, not measured.
The full version, fast.
Models and token costs are no longer the constraint on agentic work. The human who has to type each prompt and unblock each thread is. The fix is to stop launching agents one by one and start designing agentic loops, where a non human trigger (a schedule, a webhook, an outage, a customer bug report) is paired with a carefully built execution (model, skills, context, tools). Many loops running in parallel across the parts of a business become a software factory. Because those loops have to run whether or not your laptop is open, they belong on always on cloud infrastructure in isolated sandboxes. Start by automating the 5 percent of your week that repeats, then widen it.
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 – 02:19
01 · The claim
Credibility stack, the agenda for the hour, and the flat statement that programmers, editors, mathematicians and designers are all getting replaced.
02:19 – 04:49
02 · Humans, agents and the farming precedent
Where people still beat models (novel thought, intuition, practicality, taste) and why past labor shifts like farming and human computers are the right template.
04:49 – 05:49
03 · Sponsor break
PostHog read about product analytics and session replay for teams shipping fast.
05:49 – 11:00
04 · The higher level work
The ladder from doing the task, to running an agent, to designing loops, to running a factory. What the top of that ladder actually involves day to day.
11:00 – 16:40
05 · Agentic loops and removing yourself
Definition of a loop, the 5 percent on ramp, time tracking as the way to find candidates, and a live Composio query over fourteen days of calendar data.
16:40 – 19:20
06 · The mistake of assuming nothing changes
Why expecting to prompt agents the same way in twelve months is the core planning error, and the trigger plus execution split that replaces it.
19:20 – 24:00
07 · The laptop problem
Shared machine collisions, dead agents on a closed lid, non scalable hardware, and a demo of one prompt spawning 110 subagents in a three tier hierarchy.
24:00 – 30:00
08 · Local agents versus cloud agents
The Cursor adoption numbers, the security case for sandboxes, and a tour of the real scheduled automations running against his own codebase.
30:00 – 36:15
09 · Why cloud agents are hard
The four frictions that kill adoption, the open source Cloudroom pitch, the Ultra offer, and a live launch of fifteen sandboxed subagents.
36:15 – 40:10
10 · You are the bottleneck
Unused weekly allowance as a demand problem, turning tokens into revenue as the core skill, and the full roster of automations offered to copy.
40:10 – 45:20
11 · Inference economics
Six Codex plans collapsed into one Claude plan, the SemiAnalysis value chart, cross company review, and why API pricing is the wrong place to buy.
45:20 – 50:35
12 · Triggers, prompts, skills, sequences
The trigger taxonomy, a director agent that manages others, Raycast prompt snippets, the open skills repo, and a prediction that sequences come next.
50:35 – 56:30
13 · Building a factory loop live
One voice prompt creates a scheduled product gap automation with a built in adversarial reviewer, then it runs on camera.
56:30 – 1:00:56
14 · The gap, and how to see the future
Why the distance between locked in operators and everyone else keeps widening, plus the first principles thinking session he uses to forecast.
Lines worth screenshotting.
- The constraint on agentic output is no longer the model, the harness or the token price. It is the human who has to unblock every thread.
- If you are unblocking agents one at a time, you cannot keep more than about five running productively, which caps your output no matter how good the model is.
- US farming employed roughly 75 percent of American workers in 1800 and under 2 percent today, and food output went up the whole way.
- Humans keep the edge on novel thought, intuition, practicality and taste, because those come from experience that was never written down as training data.
- Agents have no sense of resource tradeoffs, so a one hour project becomes a repo flooded with tests nobody asked for.
- Running twenty agents on your own laptop is like twenty coworkers sharing one machine. They collide, they pop windows over your work, and they all die when the lid closes.
- One prompt spawned 110 cloud subagents in a three tier hierarchy: a top model as manager, mid tier coordinators, cheap workers at the bottom.
- Cursor's internal numbers went from about 10 percent of merged pull requests coming from cloud agents in January to over 60 percent by August.
- An isolated sandbox limits a compromise to one agent. A compromised laptop hands over every login, key and file you have.
- Downgrading your model subscription to save a hundred dollars is admitting you cannot turn tokens into revenue, which is the skill that decides the next decade.
- Build with one lab's model and review with the other lab's model, because different training produces genuinely different objections.
- Pick what to automate by imagining your company 100 times bigger, listing the departments it would have, then building one loop per department.
- A spoken minute is enough to create a working automation: run every two hours, find one product gap, have a rival model argue against it, archive unless the case survives.
- Flexing lines of code or counts of markdown files proves nothing, because a cheap model can generate half a million files in an afternoon.
- Sitting for 60 to 90 minutes with no phone and no computer, reasoning about one topic, beats reading more AI news for understanding where things go.
Stop starting agents, start designing their triggers.
WHAT TO LEARN
Once a model is good enough to work unattended, the limit on your output stops being intelligence and becomes how many loops you can design that never need you.
01The claim
- The list of jobs being absorbed now includes programmers, video editors, mathematicians, graphic designers and content creators, not just entry level tasks.
- Treat the replacement question as settled and spend the attention you free up on working out what the higher level work actually is.
02Humans, agents and the farming precedent
- Models are strong where training data is dense and weak at novel thought, intuition, practicality and taste, because those come from experience nobody wrote down.
- Agents have no feel for resource tradeoffs, so a project with a one hour budget becomes a repo flooded with tests you never needed.
- US farming went from roughly 75 percent of American workers in 1800 to under 2 percent today while food output rose, which is the shape of every labor shift.
04The higher level work
- The ladder runs from doing the task yourself, to running an agent on it, to designing agentic loops, to operating a factory of them.
- Higher level work means creating agents, managing fleets, splitting budget across tokens and salaries and ads, picking a model per task, building evals, and thinking.
- Anything that needs a human in the path is a bottleneck, the same way a CEO who approves every small decision caps how big the company gets.
05Agentic loops and removing yourself
- An agentic loop is a trigger plus an execution that repeats without you. A software factory is many of those loops running at once across a business.
- Do not automate a whole job. Find the 5 percent of your week that is predictable and safe, automate that, then take another 5 percent.
- You cannot automate work you have not measured, so track where your hours actually go before deciding what to hand off first.
06The mistake of assuming nothing changes
- Expecting to prompt agents in twelve months the way you prompt them today is the core planning error. Even voice prompting still spends your attention.
- Replace the human send button with schedules, webhooks, outages, error spikes, customer reports and other agents as the thing that starts work.
- Every loop is two design problems: choosing the trigger and the unit of work, then giving the agent the model, skills, context and tools to finish it.
07The laptop problem
- Running many agents on one machine is like twenty coworkers sharing a laptop. They collide, they pop windows over your work, and they die when the lid closes.
- A powerful local box only solves capacity for a fixed number of agents. You cannot buy 20 percent more machine for one unusually busy week.
- One prompt spawned 110 cloud subagents in a three tier hierarchy, with a top model as manager, mid tier coordinators, and cheap workers underneath.
08Local agents versus cloud agents
- A vendor study found cloud agents went from about 10 percent of merged pull requests in January to over 60 percent by August of the same year.
- An isolated sandbox limits a compromise to one agent, while a compromised main machine gives up every login, key and file you own.
- Real loops he runs: billing review every four hours, database health checks, customer bug triage every ten minutes, and a security sweep every 90 minutes.
09Why cloud agents are hard
- Cloud agents only get adopted when a remote thread starts as fast as a local one and brings your repos, context and skills with it.
- Most tools force their own harness and billing, which fights the fact that people already pay for the subscriptions they actually want to use.
- Pick what to automate by imagining your company 100 times bigger, listing the departments it would have, then building one loop per department.
10You are the bottleneck
- Most people finish the week with more than half their model allowance unused, which is a demand problem on your side, not a pricing problem.
- Downgrading a plan to save a hundred dollars is an admission you cannot turn tokens into revenue, which is the skill that matters most now.
- Point the strongest model at marketing, product, security and code cleanup rather than hoarding capacity for interactive chat you may never use.
11Inference economics
- He collapsed six $200 Codex plans into one $200 Claude plan after Opus 5.5 shipped, and third party analysis put the value gap near ten times.
- Build with one lab's model and review with the other lab's model, because different training produces genuinely different objections.
- Avoid metered API pricing where you can. Subscription tokens are heavily subsidized compared to paying per call for the same work.
12Triggers, prompts, skills, sequences
- A prompt template is the first rung of automation: snippets you invoke instantly instead of retyping the same instruction several times a day.
- A skill is a standard operating procedure the agent runs on one command, and it is where any repeated multi step process belongs.
- The predicted next layer is sequences, chains of skills inferred from your own agent traces when you run the same three skills back to back.
13Building a factory loop live
- One spoken minute produced a working automation: run every two hours, find a product gap against the architecture decision records, propose one improvement.
- Build adversarial review into the loop itself. A second model argues against the idea and the main agent archives it unless the case survives.
- New automations need a few tuning runs before they stop producing noise, so judge a loop by output quality, never by how many loops you have.
14The gap, and how to see the future
- The gap between people running fleets of agents and people on free chat tiers widens every month, so arriving late costs more than it used to.
- Reason forward from human behavior, market dynamics and what gets ten times cheaper, instead of reacting to whatever headline landed this week.
- Hour long sessions with no phone and no computer, done daily for a month, build a better model of the future than consuming more AI news.
Terms worth knowing.
- Agentic loop
- A unit of automated work made of a trigger and an execution that repeats on its own. Something other than a human starts it, the agent does the work, a check runs, and the cycle comes back around.
- Software factory
- Many agentic loops running in parallel across different parts of a business or codebase, coordinated so the whole thing produces output without a person starting any individual run.
- Harness
- The program that wraps a language model and gives it tools, file access, memory and a loop to work in. Claude Code, Codex and Cursor are harnesses. The model inside one can be swapped.
- Cloud agent
- An agent that runs on a remote server or virtual machine rather than on your own computer, usually inside its own isolated sandbox so it has dedicated compute and cannot touch your local files.
- Isolated sandbox
- A disposable, walled off environment created for one agent run. It stops agents from competing for the same CPU and disk, and limits the blast radius if one gets compromised.
- Trigger
- Whatever starts a unit of agent work. A schedule, an incoming webhook, a service outage, an error spike, a new row in a database, or another agent deciding work is needed.
- Skill
- A written procedure an agent can invoke by name, closer to a standard operating procedure than a prompt. It captures a multi step process you keep repeating so it runs the same way every time.
- Computer use
- A capability where a model controls a real screen, mouse and keyboard, so it can click through web consoles and fill forms that have no API.
- Adversarial review
- Deliberately asking a second model to argue against a proposal or a change. The original work only passes when it survives the objection, which filters out confident but weak output.
- Eval
- A repeatable test that measures whether an agent or model is doing the job well. Without evals you have no way to tell whether a loop is producing value or quietly producing noise.
Things they pointed at.
23:05linkimmortal-agents (Vector Labs)
46:30linkthe director (open source manager agent)
13:10toolComposio CLI
48:25toolRaycast snippets
40:20linkSemiAnalysis subscription value research
45:00linkArtificial Analysis
24:08linkCursor internal cloud agent study
Lines you could clip.
17:18
“The bottleneck are no longer the models or the tokens. Tokens are cheap.”
19:36
“Imagine joining a company and the company tells you, oh, there's 20 other people you share the laptop with.”
36:41
“You are the CEO of your agent army. You cannot be doing everything.”
38:40
“You're just admitting you don't know how to turn tokens into revenue. And that's literally the main skill of the future.”
53:15
“Don't flex lines of code. Don't flex how many markdown files you have.”
59:56
“There is no expedient to which a man will not resort to avoid the real labor of thinking.”
1:00:20
“Clicking around and doing stuff is no longer the hard part. Any agent can do that.”
24:08
“At the start of the year only about 10 percent of the pull requests that made it to production came from cloud agents. By August it was over 60.”
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.
metaphoranalogystory
My name is David Andre and over the past four years I've spent well over 2 ,000 hours coding with AI and I've also personally interviewed some of the most legit people in all of AI. Now in this video we'll cover how agents are replacing human work, a totally new way of starting agents, why agentic loops and software factories are inevitable in the future, also why cloud agents are replacing local agents and why cloud is the future, how my job and your job will look like 12 months from now and Towards the end, I'll show you live how you can build a mini software factory from scratch, and I'll actually show you how I do that myself.
So first of all, agentic work is evolving, right? It's no longer just sitting and chatting with Cloud Code 101 or with Codex. Everything changed with the release of Astra and Opus 5 .5, right?
A lot of credible people said that Astra is AGI, but whether it is or not, the evidence is here. OpenIA has just dropped this repo a few days ago. A lot of mathematicians are losing their mind because they just posted like 370 different proofs to unsolved problems that were unsolved for decades and centuries.
I'm not going to be lying to you. We are all getting replaced. Programmers, video editors, mathematicians, graphic designers, content creators.
AI can do more and more. So it begs the question, what will humans actually do? And I'm going to touch on that later in the video because it is an existential question that a lot of people are losing a lot of sleep over.
But there's actually a very clear answer and very clear path forward. It's just a lot of people haven't thought it through. Now, humans are still better than agents in a lot of things, okay?
I know it's kind of obvious, but like in a lot of ways, LLMs are more robotic. You give it an input and it does certain things, right? It doesn't generate an input.
It doesn't come with a predefined goal. It doesn't get a random thought as it's going. You can literally see its reasoning chain and it's from training data.
So let's talk about this, right? Humans are better at like creative, novel thought, at intuition and instinct because throughout... Our life, we've seen so many tokens, like audio tokens, input tokens, physical touch tokens, that are not represented in the training data at all.
Like, these things are not in text on the internet. And if it's not in the training data, the agents and the models are bad at it. Another thing is being practical.
You might notice if you have like a big task, you only have one hour to finish this big coding project, the agent doesn't know what that means. He doesn't know like, let's skip these details, or let's not fill this repo with endless tests, right? They just don't know.
like practicality and a good resource allocation. Also taste, right? This is like things like anybody can use AI to like build 1000 apps or generate hundreds of images, but only one of them will go viral, right?
Only one of the, like only a couple of them will make money. So taste and knowing what's good, what makes a great product, what makes a great YouTube thumbnail, what makes a great song. This is taste and models don't have it and the humans that do have it will be better at them, right?
So stop holding onto the things that LLMs do better. If you're still like trying to write code line by line instead of using codex or cloud code, I don't know what to tell you.
You're literally the person trying to be an elevator boy when elevators were sold, or the person doing math hand by hand doing spreadsheets when computers and calculators were invented, right? Just do the things humans are better at and let agents do the rest. Otherwise, you're going to have a very bad time.
Now, to understand where the future is headed, it's helpful to look at the past. And I think farming is a great analogy because... This is what the world looked like 200 years ago, 250 years ago.
In fact, in USA, in America, over 75 % of all workers in 1800 worked in farming, right? That's literally like vast majority of people. Today, it's under 2%, yet there is more food than ever.
Why? Thanks to technology. When I look at this image on the left, and applying for some...
document and doing some taxes some accounting on the right this is a more accurate image obviously it doesn't really represent how we will be working with agents and orchestrating them but again we're going to touch on that soon so basically the tldr is don't be scared okay stop being so afraid the past we went through multiple of these industrial revolutions human computers another thing right everybody was doing math by hand you literally had like 15 floor buildings where every single floor people are like doing math by hand humans always adopted If there's one pattern throughout the tens of thousands, hundreds of thousands of years history we have, it's that we always adapt, right?
So let's talk about it. What does the future of work look like? I call this the higher level work because if the tasks that we today are replaced, we need to know what we'll be doing.
A lot of people don't have good clarity about that. So now this section is on what the higher level of work will look like. once everything we do today is replaced.
The most common mistake I see new founders make is not enough feedback from real users. Sure, with AI, you can build software much faster, but the faster you build, the more important real customer data becomes. After all, you need to know whether you're building the right thing.
That's why I want to tell you about today's sponsor, PostHawk. PostHawk shows you how people are actually using your product, like where they get stuck, which features they're using and whether they are coming back.
And with their session replay, you can even see what the users actually did. Y Combinator, Supabase, Fireworks AI, 11 Labs, Railway, all of these companies use PostHawk. In fact, Supabase used PostHawk to spot a wave of new users coming from AI app builders.
And based on that data, pursued partnerships with those companies. As one marketing expert at Supabase put it, PostHawk has literally helped us get 10x more weekly new users than we did a year ago.
When you're shipping new features every day, your feedback loop needs to keep up. And that's exactly the problem PostHawk solves. So visit go .posthawk .com and learn how PostHawk can help you understand your users and build better products.
That's go .posthawk .com. The first link below the video. Actually, I think this graphic also illustrates it well.
Today, you just do the task. You sit in front of a computer and do a specific task. If you are watching my channel, hopefully you run agents, right?
You run Codex, CloudCode, Cursor, Pi, Hermes, whatever. So that's already better than just doing the task yourself, like interacting the agent and then do the task. This is where most people are, like literally stuck on step one or two, but we need to go to the next levels, right?
It's time. It really is time. And the next step is designing agentic loops.
And after that is running the agent factory. So as you move towards the top right in this graphic, the work is less manual, less repetitive, and more unique, more creative, more abstract. So let's get very specific, right?
In the future, you will not be starting agents one by one. This is a thing of the past. Like, yes, 2024, 2025, and even the first three quarters of 2026, this is fine.
It's no longer fine. The gap is widening. And I have a point on this later, but the gap between people who are on the true cutting edge versus people who are like kind of okay -ish using AI is going to grow.
Think of the average consumer. What are they using? Three versions of ChangeGPT, three Google AI certs.
These are using terrible models compared to Opus 5 .5 Extra High or Max. Anyways, in the future, we will be... designing things that run on a schedule.
These agents will not be started by me and you. Anything that involves a human is bottlenecked, okay? It's like a business.
A CEO cannot be involved in everything. He cannot decide when to mop the floors or when to close the books or when to like fire a junior level employee or when to start like close a marketing campaign, right? If you have a 5 ,000 person company, the CEO cannot do everything.
You have to realize soon... All of us will have a 5 ,000 entity company, just some of these will be humans and most of these will be agents, right? So you better get good management.
A lot of the things you have to realize, in AI, it's an 80 -year -old field. A lot of the breakthroughs came from mimicking biology and humans, right? The neural network is modeled after the brain.
And right now, the limit of running more agents is that you simply cannot unblock them all, especially after Opus, because it runs so fast. It's actually a challenge. If you're unblocking...
agents one by one, it's a challenge to even have more than five running at the same time. And if you sit down and realize how big of a problem that is, it becomes obvious that the human cannot be in the loop, right? In the past, it's like, okay, human in the loop and AI does something human in the loop.
We need to remove ourselves from the loop. So we need to have different triggers instead of a human sense abroad. Could be schedules, webhooks, outages, you know, product features, something behavior of your customers.
Different triggers that do not require you to type in on a keyboard or do a voice prompt. Voice prompt is better than typing, obviously, but it still requires your attention. It still requires your time and it's not enough.
It was enough maybe six, 12 months ago. It's no longer enough. So in the future, the higher level work, pay attention because this is like exactly what this video is about.
And when I talk like this, I'm actually kind of shocked. When I talked on my podcast and I always ask this question, people... they haven't thought through it.
Even these people at the frontier, they really struggle to answer what the future of work is like. And I know why, because it's not obvious. Some of these things I list out are obvious, but we don't know all of them.
Just because these are kind of clear in the fog, we don't know what's behind the fog. So based on what we know right now, let me tell you what we will do in the future. We will create new agents that are like, you know, set up for a specific thing or general agents.
Then even more higher abstract, we'll manage fleets of agents, right? So like, agent team doing your marketing, agent team doing your product, agent team doing customer support, whatever.
We'll allocate resources. So let's say you have a business that makes $50 ,000 a month. How do you spend that?
How much of that goes on tokens? How much of that goes on human salaries? How much of that goes on Facebook ads?
How much of that goes on investing into new hardware, right? We'll be allocating those resources. And also even more gradually, when it comes to agents, we'll be choosing the models.
Like when to use PyAgent. when to use Hermes, when to use Cloud Code, what model you put into these harnesses, okay? So for these types of tasks, this is a very easy task.
So maybe let's use DeepSeek V4 4 .1 Flash because it's cost efficient and we just need the best performance, right? Or this task is very important. Let's use Jeff because it's a decision model and it outputs a structured output and it cannot go wrong, okay?
For this task, we need the best of the best. Let's use Opus 5 .5 or maybe in the future, Fable 5 .5 or Astra 6 .1, whatever is in the future, okay? We'll establish evals, right?
So how do we know that our agents are performing? Are we tracking what percentage of tasks are done by agents, by humans? How many tokens are we burning?
All that stuff. Evals, benchmarks. This is what, again, higher level work looks like.
And spending more time thinking. Really, as simple as that. Thinking like, how could we take this business to the next level?
How can I use more agents? How can I create more agentic loops? What am I not doing?
This is real work, right? This is what CEOs and executives do. But now all of us will be doing it because all of our work will be done by agents.
Now let's talk about agentic loops because a lot of people are confused like what even is agentic loop? You know, there was this trend with like Peter Steinberger, like loops, graphs, ignore that completely. I'm going to be very practical, right?
Agentic loops, software factories. Basically, software factory is many agentic loops and the agentic loop. can still have a human somewhere, like at the approval process, but really a software factory should not have, right?
It should be running by itself, and it's basically the next level of that. Again, more on software factories in a bit. But what are agentics loops?
Well, it's basically the attempt to try to remove yourself. And the way to do that is just like, you know, if you watch me for a long time, two years ago, three years ago, I told you about NA10. It's not to try to automate everything like a big task.
Try with a small task, right? You need to analyze at what... work you do.
The truth is, me and you, we do the same types of tasks every day. Only when you start like a new company or like pursue a new effort or decide to write a new book, something radical, something different, that's where your work shifts. Most of the time it's like respond to the same email, open Slack, deliver to customer, you know, do another brand deal.
I don't know, it's just the same type of work, okay? It might not be the same exact thing every day, but it's the same type of thing. Which means it's an opportunity for automation, right?
Just not really possible with programmatic automation, like in the past, just plain code. But it's possible by putting an agent. And the difference now is really the model and the harnesses.
Like, massive progress on harnesses in the last year, right? Think about it. A year ago, what was the best harness?
Like, maybe cloth coat? Then, what happened? January, open cloth.
February, we got Hermes agent, right? Then, Pi agent became super popular. We got other things like Prime, another self -improving agent.
So, like... massive progress and all of these harnesses got better because most of them are open source and one of them improves and all of them just take the idea and improve, right? So that's the benefit of open source.
I wish the models stayed open source. Anyways, the point is that like progress in models and harnesses has been massive over the past 6 to 12 months and now things are possible that were just not possible in the past. Like you can just do full design, animations, insane, insane stuff, especially after Opus 5 .5.
So really there is no excuse. for you to be doing everything. Start by trying to take 5 % of your work and automating that.
The stuff that is like very predictable and safe. And then do another 5%. Look at the types of tasks what you do.
For example, what I do is I track my calendar every 15 minutes, every single day, right? So I've done that for like years now. So I have loads of data and I can literally use Compose your CLI and tell my agent.
I mean, I can just do that right now, right? Look at the use Composio CLI and tell me what I spent the last 14 days working on. Just leave out any private stuff.
I'm recording YouTube video. It's going to be seen by tens of thousands of people. So just give me like very broad buckets of what we're doing, right?
So it's going to run this. It's going to use my skill for Composio CLI. This is running inside of Cloudroom.
We're going to let it run. In about 30 seconds, it's going to be done. And then that's how you analyze.
You know, you need to know where your time is going. Just like you need to be tracking your accounting expenses to know what to optimize. It's the same thing with time.
You need to track your time in some way, maybe not every 15 minutes, some of you, to know how to optimize your time and how you're actually allocating it. Most people are clueless. They think they're like spending 20 % of time of this, turns out to be 60.
Or they think like, oh yeah, I'm somebody who works out often. Turns out they go to the gym like three times a month, right? But basically...
That's how we need to be looking at it. And in any business, in any business, the most valuable resource is the CEO's time. I don't care if it's SpaceX and Elon Musk or your three -person team.
Like my team is pretty small. Still, I'm the bottleneck on nearly everything. No matter how locked in the people on my team are.
So if you are the founder and the CEO of a business, your time is the most valuable resource and you need to know where it's going. So figure out how to wait to track it because then the next way is to replace that 5 % with a genetic loop or a mini software factory. Now, a big mistake a lot of people make with this is they assume our work with agents will be the same.
They look at like how we interacted with ChairGBT and Cloud and Perplexity and Cursor and think it's going to be the same. It's not going to be the same because we have reached the point with Astra and now Opus 5 .5 where we really are the bottleneck and it's really bad, right? All of us know, me, you, everyone watching this, we know we should be running more agents, but we aren't.
The bottleneck are no longer the models or the tokens. Tokens are cheap. The amount of inference you get from a single $200 amount of cloud subscription is insane.
If you don't have the money, like, open code go is insane value, right? Tokens are cheap, and very smart tokens are cheap, right? Especially, like, if you invest into hardware and self -host it, the tokens are free once you buy the hardware.
So it's no longer the models, it's no longer the token cost, it's no longer the harnesses. So what is missing? Everything else.
That's the problem. It's a really abstract problem. And it's what I spend most of my time like obsessing over and working on.
But you cannot be, this is literally a great visual by the way. It's like, this is how we are. These agents are waiting for us.
Like if you're running many agents, you know, 30, 40, 50 agents at once, they are waiting for us. In fact, I can just open my sidebar and look, I have so many things. And these are, okay, did you see that?
Those are my agentic loops starting. on different things right so feedback from different users of cloud room as well as like automatic agents doing cyber security and trying to simplify the code base so you know i'm just i'm not i'm not uh talking shit and i actually um do what i preach you literally saw these start by themselves i didn't do them and i'm gonna show you all of my automations later on so stay tuned but what i want to show you is like i have like hundreds i have hundreds of threads open and they're all waiting for me right except for the automations and mini software factories, which are more automated than that.
But a lot of that work is still waiting for me. And this is me literally blocking all the agents, right? Blocking the progress.
So the challenge then becomes, how do you architecture a factory, some agentic loop, something that works without you? How do you remove yourself from the loop? That is the real challenge.
So stop assuming that you will be sending the prompt, you'll be voice prompting. People who tell you that, they live in the past, okay? Again, might have been the truth a couple of months ago, six months ago, whatever.
It's no longer the truth. AI field is changing fast. You need to accept it.
If you don't like change, good luck. Things are changing faster than ever and you really have to remove your emotion from it and see the game for what it is and be willing to play the game. So the real first part, the two pieces for this is like the trigger and the execution.
Figuring out what is the trigger, right? Some of them could be on a schedule. once a day, every three hours, every one hour.
A lot of them will be on like a webhook, right? Or let's say like your database is like hitting 90 % of the connections, like concurrent connections allowed to your database. Or maybe there's a, like payments are failing and that sends a webhook.
Like that's a trigger, right? So figuring out how to trigger the unit of work and knowing what a good unit of work is and then the execution so the agent can do the same. So obviously you put in a powerful model, but then what skills you give it.
what context you give it, what MCPs and APIs you give it. That is the art, right? That is the challenge.
It's like the trigger and execution. So the nature of work will change and stop assuming the way you interact with agents today is going to be the way you interact with them 12 months from now because it will be very, very different. Also, the amount of agents is increasing, right?
Like today, in any given day, I interact with hundreds of different agents. That was not the case 12 months ago. Both me and you run more agents today than we ran 12 months ago.
That's going to be true 12 months from now. 12 months from now, we will run a lot more agents than we are today. So the challenge is figure out how to get there faster.
So you get an unfair advantage over everyone else. Now, when I try to figure out like, why aren't we running more agents, which I think is a trillion dollar, multi -trillion dollar question. Why aren't we running more agents?
Let's start with the laptop problem, right? You as a human, imagine joining a company and the company tells you, oh, there's 20 other people you share the laptop with, right? That's basically what it means running Agents like multiple agents on the same laptop sure you can do a little bit of a fix with work trees But still at a certain point agents will conflict or if you want to do Computer use the agents will do random pop -ups open random browser tabs while you the human are trying to work, right?
Also, if you close the laptop agents die if you lose Wi -Fi or internet agents die If their laptop runs out of battery, agents stop running, and that's criminal. It is really bad, especially the more agents are involved in your business and life, and the more that they do, okay?
Maybe it would not cripple you today, but again, think in the future. Once agents are doing 20%, 40%, 60 % of your business tasks, delivering on customers, researching your competitors, getting you more leads, creating new creatives and running ads for you, you know... helping you do more content, helping you improve your product, fix bugs.
If all of these automations, agentic loops, software factories stop running because your laptop is in a bad environment or just died, your business has died, okay? All it takes is for agent to like decide to open Docker desktop and run a large test suite and you already can hear the fans going. Also, if you want to be building the future where like agents are spawned by a trigger and not by you sending a prompt, You need to have an always -on server, right?
Some cloud VM, something that always is available. 99 .99, at least, with perfect internet, like 10 gigabit internet, and it's scalable. Because imagine if you buy a powerful computer.
A lot of people do this, right? They buy a powerful Linux box or powerful Mac Studio, and they think they've solved the problem. Well, okay, you only solved the problem for X amount of agents.
What if you need that amount plus 20 %? You cannot just upgrade. Your existing box on your computer and at 20 % you have to buy another box, right?
And what if you only need a 20 % for a couple of days and for a week? so like you need this to be dynamic scalable both up and out and Let's say if you want to launch hundreds of agents most of you cannot do it Most of you just cannot do it on a computer because your hardware is not powerful enough But what if there was an opportunity for you to do that?
What if that would be a big customer that needed to do that work or something like you know opportunity in the stock market, whatever. Just something where you need to run hundreds of agents at once.
But you cannot do that because your computer is limited. Let me show you how we did it. One of our engineers at Cloudroom literally just solved that problem.
Or I tried to split us out because people still live in the old paradigm of launching agents one at a time. But you can simply tell... In Cloudroom, one agent to start 110 agents, right?
And this was a three -tier hierarchy. So the most powerful manager, like the CEO, whatever, was Opus. Then all of the coordinators, managers were Sonnet.
And the workers were Haikus, right? And this, again, let me just make that very clear. Simple prompt.
This is how you need to think. It's like, what if you design one agent that can start 10? You see the leverage on that?
Instead of like launching all of the workers one by one, think of a task where... 20, 50, 100 agents can actually do meaningful work. And think of like what you need to do to ensure they do good work and they just don't do a much of slop, right?
That is the challenge of the future is figuring out the skills, the rules, the conditions, the user interfaces, the review processes, all of that stuff, right? And the first step as I began working on this, like why are we running more agents? It became obvious that it's a laptop problem.
Now, I try to build around this, right? I actually open source in multiple repos. So I did like, it's called immortal agents.
And basically, anytime an agent goes to sleep, it would start it. Okay, there it is. I found the repo.
It's under our Vector Labs org. But basically, anytime the agent goes to sleep, it gets restarted, right? So any agent that would die from loss of internet or you closing the laptop, when you reopen it...
it would start itself. That way, your agents are running. Your agents keep running.
Obviously, this is more of a tool, but that prompted me to go in this right direction of why aren't agents more running? And a big problem with that was the agents running local. Let's talk about this.
Local agents versus cloud agents. A lot of people don't know the difference, but in fact, Cursor did a study internally where at the start of the year, only like 10 % of their agents, 10 % of the PRs that made it to production came from... cloud agents by august it was over 60 so from 10 to 60 in a matter of you know eight months or whatever if you have good visualization skills you can imagine you know a scale where this is 10 in january and this you know january and this is august at 60 here right so it's not doesn't take a genius to see the exponential growth of that and soon enough it'll be 90 probably won't reach 100 Because there is still some value in running sometimes agent locally, like, you know, for privacy or for like cleaning up your folders, whatever, right?
But the vast majority of tasks, I would say over 95 % for sure, probably over 98%, can be done better in the cloud. So what is the difference? Well, a local agent runs on your laptop or your PC, whatever.
Cloud agents runs on a server, on a VM, isolated sandbox. Your laptop or whatever computer you have, even if you have the top, you know, DJI workstation, top Mac studio, it still has limited amount of CPU, RAM, and disk, right, that you purchase and you're stuck with it.
Cloud, if you, especially when you put each agent in an isolated sandbox, you're only paying for when the agents are running, and it's infinitely scalable. You can have 10 agents, 10 sandboxes. You have 500 agents, you have 500 sandboxes.
You never hit the limit, and this is like one of the first things. You need the environment to be correct. There's a lot of other factors to making sure your agents are always running and that they're not limited and blocked by you.
But the first thing is making sure their environment is correct. And if you run them on your laptop, that is going to be limited in a lot of different ways. Also, computer use.
This is very annoying. Computer use is something that has been massive progress in the last three or four months. If you haven't ever tried like Codex computer use or Cua driver, you are missing out.
The agents are amazing. They can go and take over your screen and click around and fill out forms and set up AWS. Google Cloud Console and the most convoluted interfaces known to man, you never have to use them ever again.
You just give a powerful model computer use and that's it. Problem is, they take over your screen. They do pop -ups, you know, ask for permissions.
They open new browser tabs. You are trying to work. It's good if you don't want to work on that computer, but you're trying to work on there, right?
So that's another thing. It's very inconvenient having 20 plus agents running while you are working on the same machine. Another thing is cybersecurity, which is becoming more important with each passing day.
In fact, I recently interviewed the CEO of Vercel, Guillermo, and he said, quote, cloud is infinitely better for security. And it makes sense, right? If you have your own device, I mean, you can just draw it real quick.
Boom, this is your laptop. And then you have a bunch of sandboxes where the agents are running in the cloud, you know. If one of these gets compromised, at worst, that agent is compromised, right?
If your main machine gets compromised because you were running some bad model and somebody injected it and told it to download a malicious package or something, then you lose everything that's on your laptop, all the logins, all the data, whatever. But if one of these sandboxes gets hacked, it's just that sandbox. So as cybersecurity is more important, both for individuals and companies, the more the value is running agents in cloud and specifically in isolated...
cloud sandboxes i already showed you how automation started on its own actually i want to show it now but you saw it earlier where inside of cloud room i literally have these automations and uh i mean i can just explain some of them right so for example billing review checks all the billing code every four hours um to see if there's like any issues right any new bugs anything i missed last time database health check this can change all the time right suddenly maybe multiple power users wake up and use it at the same time and maybe like again the concurrent connections don't keep up or CPU is not keeping up or or something or like we're running out of memory right there can be many reasons many things that go wrong it's good to have the agent investing and actually this is a software project right so all of these would be departments in a large software company if you thought about 5 000 software company you would have a team database team you would have a team just on the billing and and payment right you have team on like
cybersecurity and authentication these would be teams in a large software company so whatever you do think of like how that would split up if your company was 10x bigger how would that split up into teams and then think of like the agentic loops you can create today for those different parts of your business right this one is especially good anybody who has a bug inside of cloud room can literally just tell report this to david and This automation runs every 10 minutes, checks a table inside of Supabase, and if there is a new entry, it spawns Opus 5 .5 agent to investigate it.
Now, the beauty is that this doesn't take my time. It's fully automated when it's like something small, right? Like some rare edge case or like an idea that doesn't really align.
If it's a serious issue or if it's like a great feature idea or something that is worth fixing, it investigates it, explains it to me in a couple of plain English sentences. and leaves it be. But if it's not, then it's self -archives, right?
So whatever threads you see here, these are the things that it decided to keep. All of the stuff that is like, that these autonomous agentic automations started and was small or like there was no issue. Maybe, again, checks the database every five hours and most of the time there is no issue.
So Opus 5 .5, the smartest model in the world, runs on like extra high or max and investigates something and self -closes if there's no issue. Like, that is tremendous value. If there is an issue, it spots it, explains it to me in plain English, and tells me what the proper fix is.
Again, in the future, this will be even more automated. Even more of these will be launched by agents. But you can start to see how powerful this is.
Me, as somebody who's running, like, a very lean team, we are doing so much more work than literally companies with, like, 50, 100, 200 developers. Why? Because we know how to use these agents, and we have these agentic automations, and...
You saw earlier in the video how multiple of them started at the same time. And by the way, stuff like this is absolutely essential. Security sweep, right?
Every 90 minutes, just extensive pen testing and cybersecurity testing of my code base. If you are not doing that, someone else will do it on you, right? So in the future, it'll just be whoever spends more tokens on a software and tries to, you know, break through, tries to access to sensitive data, get admin access or, you know, DDoS you, whatever.
So this is not really optional. Like, all of you will implement this. The question is...
will you start implementing it today or will you wait six months, 12 months and implement it once it's obvious to everybody, at which point there is no advantage. Now, cloud agents are hard, right? There's a reason why most people still don't use them.
And the best products are all closed source from large companies like Codex, Cursor, Cloud, Devin from Cognition AI, right? All of them are closed source and there has never been a good solution that lets you like... to have full sovereignty over data, export the chats, and that allows you to use any model, any harness, any subscription.
And there's a reason for that, because cloud agents are hard, right? It's hard to start a thread and it needs to feel as local. If it doesn't feel just as fast as on local, people will not adopt cloud agents.
You need your repos there and your context there, your skills, right? You want to use the subscriptions you already pay for. If you need to switch to like, I don't know, amp and devin the problem with them is that they force you to use their subscriptions and their harness but all all of us are paying more for ai tokens than ever before right we want to use the subscriptions we already have whether that's codecs cloud code cursor news portal whatever and the agent sometimes still needs a laptop so cloud agents are hard it takes a lot of care and effort and deep knowledge to make them execute correctly and if just one of these frictions would be enough for people to not use cloud agents.
So that's exactly what my vision is with Cloudroom, what I'm building. This is a platform for running thousands of agents productively. That's the ultimate vision.
And I'm building it open source. And it allows you to use any model, any harness, and any subscription. And it's really designed to not only do the current work, like I showed you earlier, with chatting with any model, whether it's cloud code, cursor, codecs, this is Devin, Pi, OpenCode, Hermes, whatever, but also to do the futuristic steps, right?
Creating the agentic loops, agentic automations, moving towards the future of work with the software factories and having all the tools where you could run thousands of agents productively from a single interface and also letting you choose whether you start thread on local or on cloud because, you know, sometimes you really might want to clean up your downloads folder or desktop and you want to run it locally or local in a work tree or just in a cloud where, Each cloud thread gets its own isolated sandbox.
So the target avatar for this is really not everyday people. It's entrepreneurs and really the superpower users, advanced developers and people doing a lot of agentic work. And if you aren't already paying for multiple $200 a month subscriptions, this isn't like, you're not the target avatar for this.
The ideal person would already have $200 plan on Cloud Code and $200 a month plan on Codex. And you already... should have the need to run agents in the cloud like my job here is not to convince you that you should be running cloud agents right you've already watched enough of this video if it's not obvious by now then i probably am not gonna convince you right so the reason i'm saying this is because we're moving towards general launch and before that i'm looking to work with a couple of individuals a handful of individuals who have a business or who have the need to run many many agents on cloud infrastructure Because I'm opening up Cloudroom Ultra.
So when you go to cloudroom .dev, which I'm going to link this below the video, you can click on get started or just the pricing. And Cloudroom Ultra is really for the most powerful people. And again, I'm going to be working with you directly to help you deploy cloud agent infrastructure for you.
You're going to get personalized onboarding with me and my team. And you get 10 ,000 cloud agents for 30 days. So 10 ,000 different cloud agent launches.
And each agent runs in its own environment. So again. I'm going to demonstrate that to you.
When you are inside of Cloudroom, you can select the environment. So here I can say, launch 15 sub -agents to analyze different parts of this code base for small bucks. Put each sub -agent as Sonnet 5 .5 extra high model.
Boom. So here it is. It launched the main agent.
And look at that. That was like less than a second, okay? Most cloud agent platforms take 20 to 30 seconds to boot up the sandbox.
We have done a lot of work, me and my team, to make this feel as fast as possible. As fast as local. That really is the bar.
It needs to feel as fast as running agent locally. So this Opus will launch 15 sub -agents, and it's analyzing the codebase, and it's going to make each of them look for different small bugs. And it could be any other task, obviously.
This is just a demonstration task. It could be like doing research or cybersecurity testing or working on a feature, and they're this. Look in the left sidebar.
It's launching these. very easily and this is already something that like cloud desktop cannot do it cannot like interact with other threads programmatically like that and the beauty of the cloud threads is that each agent run its own isolated sandbox so they're not stealing resources from one another they are not uh competing for the same cpu and ram and disk right let me just stop it now pause all of them you can also see all your limits right from all the different plants boom and just like that he stopped all of them Now archive all of them and reload them as codecs instead with GPT 6 .1 Sol as the model and do 25 subagents instead of 15.
So if you look at here, it's obviously Sonnet because that's what we discussed. But now, boom, it archived all of them just like that. And now we should launch 25 codecs subagents with a different model, right?
So you can mix the harnesses. You can have Opus as orchestrator and then Sol as the worker or Astra as the manager and then Opus or Sonnet as the implementing agent. It's entirely up to you, right?
All of this is possible inside of Cloudroom, and that's what I'm trying to build, is the platform to run thousands of agents productively. So as I mentioned, right now I'm looking for a handful of people to work closely with, people who can afford the Ultra plan. And again, your target avatar is somebody who has a business, who has the need to run 20, 30, 40 agents in parallel.
Someone who already is paying for $200 a month, Cloud Code and Codex, and wants really to be at the cutting edge of... and it already sees the value of cloud agents. So if that is you, go to cloudroom .dev and join the Ultra plan.
And again, all of you will get a personalized onboarding from me. Let's talk about this principle that you are the bottleneck. This is essential to understand the future of work because if you don't remove yourself from more and more things, you cannot scale.
And it's the same as a business, right? If you want to scale to like 10, 20, 30 people, you need to start removing yourself from some tasks. You cannot be doing everything as a CEO.
And the same thing is... . You cannot be doing everything.
So Opus 5 .5 is already AGI level. If you disagree, you are wrong. It's the best model right now.
Buy a lot because it's extremely cost efficient. I'm going to touch on like what is my token setup and subscription setup a bit later. But Opus 5 .5 is like incredibly fast and incredibly token efficient.
And it's incredibly smart. All of the design and visuals, like literally I'm going to admit it, right? If we go to Cloudroom, all these animations and these graphics.
This is me working with Opus 5 .5, right? Yes, I asked for multiple variations and I did some tweaks to make these animations smooth, but like no other model would come even close. It's crazy.
Like all of these designs, yes, you need to have some taste as a human to know what is good and what isn't. But man, like Opus is AGI. It's AGI.
So you need to figure out how to let it cook, how you can get the most out of the subscription and remove yourself. That's another thing. A lot of you...
Let the weekly limit just go with like more than half of it unused. You can do a lot with that. And the bad answer, there's two answers, right?
One of them is like, okay, you're going to downgrade your subscription. Sure, you save $100 a month. But then you're just admitting you don't know how to turn tokens into revenue.
And that's a very bad problem because that's literally like the main skill of the future. Can you use these tokens to generate more money than it costs? cost you to pay for the subscription.
And the Anthropic subscription is already subsidized, like, 90%. And more on that later. I'm going to show you the actual stats, the research from semi -analysis.
But, like, you get, like, 90 tokens of Opus 5 .5 for, like, 90%, 50 % to 90 % less than you would get it in the API. So what do you do? Well, you try to do something, right?
You try to improve your marketing, try to improve your product, cybersecurity. I already showed you, like, guys, just steal it. Steal my automations, right?
billing review, database health check, feedback threats from customers, GitHub issues, PRs, overview of like public repos, sandbox alert threats, different issues with sandboxes and like hitting limits, security sweep, cybersecurity, simplifying the code base, looking for unnecessary complexity or unused code, right? We can go a bit deeper here.
Test gap. Okay, so this one is like a bit more advanced. This one looks for missing tests.
but it's more nuanced because agents love tests they will just flood your whole code base with unit tests and component tests integration tests whatever this is much better because it checks different parts of the code base to see if there's a like a gap for obviously good test when opus thinks of it it launches gbd 6 .1 sole extra high doing adversarial review of it basically kind of assuming that like is this test you know assume it's not needed whatever whatever and Only if both models, Opus who found the test and GBD who was reviewing it, if both agree that this is a good test to add, then it passes and then it goes to me for approval, right?
So you can see like how these agentic automations can become more advanced over time. And really you need to start building these for you. So whatever that is, whatever, if you're building a software, if you're doing a marketing agency, if you're a real estate agent.
You need to figure out if your business was hundred times bigger and you can just research it because there probably are companies hundred times bigger What are their departments like what are the teams at their companies like and that's how you need to start creating these automations that again Some of them run on the trigger some of them run every ten minutes some of them run every four hours Depends up to you and that's like the really the art of you fine -tuning that over time So basically the question is what else can I point Opus at or Astra or GPT 6 .0 these super powerful models How can you use them?
What is like some automation, some loop, some mini software factory you can launch that would like be directionally correct? And again, you need to try some of them. Maybe you delete some that are not useful, but if you don't try, you will not find the great ones.
Okay, so let's talk about inference because this is one of the main variables of running your software factories, your agentic loops, whatever. It's like how much you're spending on inference and what tokens you're running. So great research from Seminalysis, which by the way, I posted like a less scientific version of it.
on Twitter and it went viral recently. Here it is, a lot of people thought I wasn't, I was being like, fastidious or like, I was being overly dramatic, but this is actually true. Before Opus 5 .5 released, I had six different $200 a month Codex plans.
Why? Because Astra was the best model in the world. For a solid three weeks, Astra was the GOAT.
It was by far the best model in the world, way better than Fable 5 .1. But then Opus 5 .5 released and won! Cloud code subscription for 200 can handle the same amount of tokens in fact more because Opus is insanely fast way faster than Astra and That's why I was very confident in this Estimate that currently the cloud code 200 subscription gives you 16 to 20x time or usage Then the codex 200 and then obviously open AI before the dev data dropped They reduced the usage of codex 200 plan by half and then they released their 500 dollar plan.
So You know, it is what it is. But basically, semi -analysis, even though it wasn't as radical, kind of confirmed that it is like 10 times more value, 10 .6 and 10 .6.
So basically, you get 10 times more value currently from Cloud Code subscription compared to OpenAI Codex. So if you only have like space for one $200 amount of plan, it's not even close. I cannot in good faith recommend you the $200 Corex plan right now when the current Anthropic plan is such a good value.
So Opus 5 .5 is the best model if you think otherwise you're wrong. And again, there's probably going to be another big model release in the next two weeks, three weeks, who knows. So this can change very fast.
But Anthropic is right now on a tear. And I expect if they release Fable 5 .5, probably going to be the best as well. So now, basically, most of my tokens come from a single...
$200 Cloud Code subscription. Now, I did get a second one to run more of these agentic loops. So to be fully transparent, I'm paying for two.
And yes, I'm paying for all of these. The companies are not paying me for them. It's my own money.
And this was worth it at the time. Astro had the best tokens. So I was happily paying $1 ,200 a month for six different Codex plans.
But now Opus is best. So GPT models, how do you use them? Well, you use them for review, right?
So if we go into Cloud Room. What I might do is for example actually here is like an example of automation how it works when it runs, right? So this was a simple like a 16th execution of this automation to simplify the code base and it's it comes with the prompt launches it and then It ran for four minutes 12 seconds and this is the model it was sold with max and it found a cleanup So now what I can do is actually usually we in the reverse so say run opens review To see if this is legit.
And save to delete. Right. So he will obviously run this as a sub -agent.
And again. This would be inverted. You always want to review with the separate.
Company model. So Frontier are two companies right now. OpenAI and Anthropic.
Other companies are trying to be Frontier. But it's not even close. OpenAI and Anthropic are the two companies.
At the cutting edge. So when you build something with. GPT 6 .1 Sol or GPT Astra, 6 Astra, you want to review it with Opus 5 .5.
If you build with Opus 5 .5, you want to review it with GPT 6 .1 Sol, right? Always review with the second company model to spot like differences and to look at it in a different way. So again, 98 % of my threads are running Opus as the primary.
So then I launch GPT 6 .1 Sol or Astra or sometimes both if it's like a serious change for review. Then I ask Opus, like, is this legit or is this just overthinking, right? You want to avoid API pricing at all.
Don't pay API pricing. Yeah, the subscriptions are a better deal. If you need to pay API financing for some cost, like some reason, just you would need to use like Chinese models for sure because they're a much better deal.
I mean, now Haiku is pretty good deal from Anthropic. And yeah, you can look at these more specific numbers from artificial analysis. Anthropic is just giving you a much better deal right now, and it's not even close.
So, let's talk about the types of triggers, right? Because this is what matters when it comes to building these software factories or agentic loops. What starts an agent?
In the past, or I guess currently as well, most of us are still launching a lot of agents manually, which is fine for a transitionary period, right? It's fine to do it where it's nearly needed, but you should really seek... to remove yourself from as many of these as possible and look for different types of triggers such as a schedule every 15 minutes every five hours once a day webhook this is going to be one of the most like flexible ones a separate like incident database outage error spike whatever customer feedback right the more customers you have the more feedback you have proactive uh bug hunt which uh can run on the schedule i guess but or another agent that's another thing right this is going to be really the frontier for example one of my open source repos is the director let me see if i can find it here and that's a like a main agent yeah right here that manages other ones and that kind of learns your prompting style and learns your decision making over time and again this is an open source repo so you can take it but really this is going to be the hardest frontier and that's where it's going to get the most crazy because you can imagine like i show you you can have a manager
with 15, 20, 25 sub -agents and if you have good skills and good knowledge and Good rules how you operate and you can teach that to a powerful model like Opus 5 .5 as a director You can get probably very far and it can actually unblock the other agents more efficiently than you would think So this is important to understand and again triggers are like half of it.
The second half is the execution Now, let's talk about prompts and skills, right? Because people don't see what that is.
They think, oh, it's just a markdown file, but really, this is the first attempt towards automating and towards optimizing. A skill is like something you do a lot, right? Prompt is like the simplest.
A prompt template is just something you do all the time, right? I can show you a bunch of these that I use every single day. Like, for example, this.
Rebase off of main. Check the underlying code to see if it has been fixed or not yet. Answer in short plain English.
Or... answer in short in plain english or don't be verbose make your answer simpler and shorter focus on what really matters or run 4x varied search 3x deep research and 3x deep scrape on this or anything else to do here or not really remind me of the original purpose of this thread these are all raycast snippets that i can just instantly invoke faster than you can even understand what's happening and uh yeah that's how i can work way faster than anybody else but these are still like prompt templates right the next level of that is skills i have many skills obviously in fact my skills repo is open source so you can just take them and copy them the ones that matter and uh yes this is my most popular repo over 4000 stars and skills are basically like more advanced um processes right like sops and companies standard operating procedures things that you cannot repeat and find yourself repeating such as the decision skills that i run after it does or like the gpt reviewer you know like running it with the
with the opposite model like the review skills i have like a total review as well or i have like a gui release or types of different releases right a core release computer use skills um yeah a lot of different stuff so the types of processes like the types of work you can turn into skills and then you can invoke them but if you're paying attention in this video you are still needed in that loop and that's not ideal all right so prompts and skills are a great step in the right direction it's definitely directionally correct but you still want to be removing it and you still want to do like if you need to be involved more work should happen so what i predict is that after skills we will see the invention of sequences and sequences will be basically a sequence of multiple skills chained together but it's going to be based on probability of watching your own behavior right so all of us have agent traces and like the agentic sessions on our computer and
if you've been using agents for a couple of months you probably have thousands of different agent sessions stacked up so somebody could see like okay in this type of situation david runs these three skills back to back with a 98 probability that would be a sequence a sequence of skills right and uh i think that's gonna be the next step in the right direction again you still need to be involved to some degree So we're going to be optimizing both like asynchronous work and synchronous from the human.
Both of these need to be hyper -optimized. But after skills, I predict that sequences will be the next big paradigm, big strength over the next six months. Now, agentic loop into software factory, right?
So loop is the trigger where the agent doesn't work and there is a check and it just repeats, right? A simple loop. Factory is many different loops doing different things, right?
So I showed you like, I call it a mini software factory because... I don't think we're quite there yet. I think this is still early days, so I don't want to like fully claim, like a lot of people just say like AGI, they were speaking it for last three years, like AGI is here, whatever, software factories.
I want to make the claim when it's there, okay? So I think we have reached AGI, but I don't think people are running, like I wouldn't claim what I have right here is the best software factory possible. That's why I call it a mini software factory, right?
So it's not just one agentic loop. You can see it's like on different... like software disciplines and types of software work, because I'm building a software company.
Obviously, if you're building a different type of company, it would look different, but that's why I like to call it a mini software factory. And again, software factory is just many different agentic loops working together in parallel to achieve a specific outcome, either in your life or business. So yeah, that's the difference.
So I guess let me show you how easy it is to build an automation in plain English inside of Cloudroom. And I already have many so let's uh Let's do a new thread. Let's do Let's do a cloud thread.
Let me show you how fast cloud threads are Opus 5 .5 high is enough. I'm gonna voice prompt it, right? I want to create a new automation and basically Actually, let me show you a local thread because I already showed you cloud threads.
So I want to make you See all of it that you can do stuff on local as well. Okay, if you're using codecs or a cloud code or whatever you can do local work inside of cloud room as well so let me show you local because i already showed you how powerful it can be on cloud running all of these uh subagents so let's do this create a new cloud room automation where every two hours a new thread spawns with opus 5 .5 extra high as the model and it looks for product gaps against all of the adrs we have in slash docs slash adr and looks for like one obvious improvement for our product.
It could be a feature, it could be more clear UI, anything. It just needs to be aligned with our ADRs and decisions we've made in slash docs folder. And then when it finishes, and it has a good idea, launch a reviewer, a sub -agent below it, with GPT 6 .1 fast extra high as the model.
and doing adversarial review to try to make the arguments of like why this shouldn't be a feature and why this shouldn't be added. Once the sub -agent finishes it should go back to the main agent and he should review the feedback of the sub -agent thinking about it from first principles and if he agrees with the sub -agent he should simply archive both threads but if he disagrees he should keep this thread open waiting for me, David, to review it.
Create this Cloudroom automation now, keep it simple, don't overthink it. So just like that literally less than a minute of speaking and i have a prompt that will result in an automation that can like forever improve my product right now again when you launch a new automation it might need a few like tweaks in the first time first few times it runs to get it right to make it useful so that it's actually producing good results and not just wasting tokens or you know adding slop this is the main thing a lot of people just flex like Yo, recently I literally heard somebody flex that he has over half a million markdown files.
Like, nobody cares. Like, you could just run a cheap Opitus model and generate half a million markdown files. Like, literally nobody cares.
So, don't flex lines of code. Don't flex how many markdown files you have. It just shows you're a dummy and a beginner.
You need to figure out how to add this in a way where, again, you... I think big product decisions should be made by humans. And, like, UI UX still...
Agent does have that taste. But again, back -end bugs, database optimizations, speed -ups in the codebase. A lot of these things.
Opus is much better than me. And I have no ego around that. I have no need to hold on to things where Opus or Astra or other models are better than me.
So you need to let go where the agents are better. And instead, embrace it and build those engine loops and automations. This is the test run.
Okay, so it's looking for it. It's analyzing the codebase for an idea.
and we'll see if the sub -agent launches with Sol. That's how it looks like, product gapped automation. And by the way, the nice thing that you can just click through them and these agents are tagged.
So you can tag any of these agents. So for example, here I can do like find bugs across 15 codes areas, just archive this thread and all of his sub -agents. And these agents can interact with each other.
You can have like complex relationships and hierarchical structures. Really, your creativity is on limit. And my goal with Cloudroom, there it is, all of them are archived.
My goal with Cloudroom is to like build the best open source cloud agents. And really right now, we're the world's first tool, like AI coding tool that gives you a cloud agent in each thread with a separated environment, right? So here, by the way, you can say like, What computer are we on or what machine are we on?
Gonna run that same prompt in a new cloud threads so you can see The difference right so here it'll say like over on MacBook Pro, whatever whatever 128 gigabytes of RAM There we go. Right then the same question here. Look how fast it was.
It was the same speed as local We're in a cloud room cloud sandbox. So yeah, this is a This is my goal is to build the ultimate platform to run thousands of agents proactively.
And yeah, there it is. It launched the... It launched the subagent to do the adversarial review.
To see like what's happening and whether this is a good feature. So yeah, that's how easy it is to launch a new automation. You can actually see it here.
We have now 32 instead of 31. And I think this is that one. Product gap every two hours.
So yeah. I mean, you can turn it on, turn it off. Very easy to manage that.
So, another thing before we wrap up, I want to address a complete lie that you might have heard from people like Sam Altman, and it's that AI will benefit everybody and it's going to lift everyone. Guys, the gap is growing. Right now, I'm in San Francisco and I met so many Lockton people, and everybody agrees that if you know how to run 50, 100, 200 agents productively in a useful way, it's not going to be even fair compared to the average person on the street.
And again, most people still don't use any AI, right? Like most people, your friends, family, whatever, they're not locked in at all. But even those that do use AI, they're using a free plan on ChaiGPT, a free plan on Gemini.
They're not building software factories. Like the gap is growing. The people who are locked in, who are at the cutting edge of AI, they're getting further and further ahead and they're getting like insanely more productive compared to everyone else.
So really the cost of not being on the cutting edge has never been higher. And yeah, the reason why like these public CEOs of big companies, they don't say it is because, you know, they're on a big interview on TV or whatever. And it would be really bad PR if they said this.
But I don't care. I'm going to say the truth. And the gap is growing.
So you might be like, okay, David, how do I see the future? How do I reason about it, right? The thing is like a lot of these things are kind of obvious if you just sit down and think about them.
What we do at our team is just like we do a one hour session and like, you know, you can literally do like a voice recording on your phone or on your computer and we reason about a specific topic, right? Like why is this company growing? Or we notice this trend behavior of people with AI.
What does that imply? Or when Jeff released, like what do like cheap decision models imply, right? And you just sit down and you don't do anything.
No phones, no computers. You just reason and talk. on a specific direction for 60 to 90 minutes.
And if you do that every day for a month, for two months, you will have a better understanding of the future than even most people in AI, most people on the cutting edge. Because again, thinking is hard, right? So you need to reason from first principles about human behavior, market dynamics, new technological breakthroughs.
What will change? Like if there's something that happens, what are the second order consequences, third order consequences? And really reason through it, and don't leave the table until ideas start flowing.
Usually, the first 5 or 10 minutes are not productive, and it's kind of stuck, and there's nothing unique. But after that, you have ideas. So that's also how I reasoned my way into Cloud Room, and that cloud agents are inevitable.
It's just by reasoning through it, and realizing what is going to be the bottleneck, and realizing that the models are already strong enough to run hundreds and thousands of agents, but we are not running them. Most of us are not doing that. having thousands of agents productively.
So think about what is the bottlenecks and trying to solve them once a better time. And really you have to allocate time to this because as Joshua Reynolds said, there is no expedient to which a man will not resort to avoid the real labor of thinking. In other words, people will do anything else still.
They'll happily work at McDonald's 10 hours a day for 40 years just to avoid sitting down and thinking like, how do I get a better job? Or how do I improve my career? Or maybe which industry should I work in?
Thinking is the hard part, okay? Clicking around and doing stuff is no longer the hard part. Any agent can do that.
You give an agent a computer use and a couple of MCPs and tools, they can do a lot of things better than humans. The real labor, the real work, difficult work right now is going to be thinking and reasoning about the future from first principles. So make sure you actually do that and make sure you're not avoiding the painful work of thinking.
So with that being said, thank you for watching. And again, if you're interested in working with me more closely, I'm going to be working with roughly the first 10 people who sign up on the Ultra plan for Cloudroom. I'm going to personally onboard them to the platform and help them really adopt cloud agents in their life and their company.
And yeah, so if that's you, if you're a business owner, if you're a real power user that runs at least two $200 plans, whether it's on Codex or CloudCode, you are the target avatar. if that's not you no worries this is uh this is not yet for you but without being said thank you guys for watching and hope you find this video valuable and have a great productive week
The Hook
The bait, then the rug-pull.
The opening does not sell a tool. It sells a deadline. Two thousand hours of agent work and a wall of interviews get spent on one claim: everyone watching is getting replaced, including the people who already use AI well, and the only question left is what the higher level work turns out to be.
Named ideas worth stealing.
06:12model
The four levels of agentic work
- Do the task yourself
- Run an agent on the task
- Design agentic loops
- Run the agent factory
A ladder of abstraction. Each step up makes the work less manual and repetitive and more creative and unique. Most people are stuck on level one or two, interacting with an agent for each task.
Steal forauditing where any repeated workflow currently sits and what the next level up would look like
06:28list
What the higher level work actually is
- Create new agents
- Manage fleets of agents
- Allocate resources
- Choose the models and harnesses
- Establish evals
- Spend more time thinking
The six activities that replace task level work once agents handle execution. Resource allocation means deciding how revenue splits across tokens, salaries, ads and hardware. Model choice means matching a cheap model to easy tasks and the best model to high stakes ones.
Steal forwriting a job description for yourself twelve months out
18:35model
Trigger plus execution
- Trigger: schedule, webhook, outage, error spike, customer report, another agent
- Execution: model, skills, context, MCPs and APIs
Every agentic loop breaks into two design problems. Half the work is choosing what starts the unit of work and what a good unit of work is. The other half is equipping the agent to finish it without supervision.
Steal forspecifying any automation before you build it
21:30model
The three tier agent hierarchy
- Top model as manager
- Mid tier models as coordinators
- Cheap models as workers
Instead of launching workers one by one, design one agent whose job is to start ten others. A single prompt produced 110 subagents on this pattern, with the expensive model only used at the top.
Steal forany task where fifty parallel passes beat one careful pass
27:00concept
The 100x company method
To find automation candidates, imagine your business a hundred times bigger and research what departments companies that size actually have. Each department becomes a candidate loop you can build today, running on its own schedule.
Steal forturning a vague feeling that you should automate more into a concrete list
49:00model
Prompt to skill to sequence to loop to factory
- Prompt template
- Skill
- Sequence of skills
- Agentic loop
- Software factory
The progression of automation maturity. Prompt templates remove retyping, skills capture multi step procedures, sequences chain skills based on observed behavior, loops remove the human trigger, and factories run many loops at once.
Steal forworking out which rung your own setup is actually on
44:10concept
Cross company review
Whatever one lab's model builds, have the other lab's model review. The argument is that different training produces different blind spots, so the second opinion catches things a same family reviewer would wave through. Only act when both agree.
Steal forany change important enough that a confident wrong answer would cost you
59:05concept
The first principles thinking session
Sixty to ninety minutes on one topic with no phone and no computer, reasoning out loud about human behavior, market dynamics, what gets ten times cheaper and what stays scarce. The first five or ten minutes produce nothing. Done daily for a month or two it builds a better forecast than consuming news.
Steal forany strategic decision where the answer is not in a dashboard
How they asked for the click.
VERBAL ASK
35:00product
“So if that is you, go to cloudroom.dev and join the Ultra plan. And again, all of you will get a personalized onboarding from me.”
Qualified rather than broad. He names the target (a business owner already paying for two $200 plans who needs 20 to 40 parallel agents) and explicitly tells everyone else they are not the buyer yet, then repeats the same offer at the sign off.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
