
AI agents already get jobs and money. Where this leads in a few years and how to prepare
The news about AI agents says opposite things every day: one day they are taking jobs, the next day the people fired because of them are being called back. If you are building a project with an agent yourself, it is hard to know what to count on a year or three years from now. From this episode you will take away a simple way to make sense of such news: look not at who promises what, but at what has been measured and how that number changed over a year. I talk about what is already happening with agents and where, in my view, it will lead in two to four years. You will learn what work an agent already does as well as an ordinary adult and where you cannot rely on it, how agents are allowed to pay today and why some large companies keep them out, when a strong model will run on your own inexpensive hardware, and how things really stand with robots. Each time I say what is an established fact and what is my forecast. At the end, six short pieces of advice. How to give an agent access to money so that it cannot spend more than allowed, even if someone tries to trick it. Which words to remove from your instructions to the agent. How to cut work into pieces and how to check the result.
Transcript
Tap a paragraph to listen from there.
Good evening, startup founders and vibe coders! Today I am going to tell you how AI agents are becoming independent players in the economy right now: they are getting work, money, their own computing power and, little by little, a body.
This episode has one idea, and I will build the case for it step by step. In a few years, an ordinary agent may control more money than an ordinary person. I know how that sounds. So along the way I will say plainly what is a verified fact, what is a trend we can measure, and what is my opinion. And closer to the end I will tell you what this means for you, since you are already building your projects with agents.
I will start not with agents but with ourselves. With how people judge something new. I already know the main objection to my idea: come on, this is science fiction. So let us talk about science fiction.
Think back to how the first language models were received. In the summer of 2020, the scientists Gary Marcus and Ernest Davis wrote a column about GPT-3, an ancestor of the models you work with today. They called it "a fluent spouter of bullshit" and wrote that "it's not a reliable interpreter of the world." And they were not being unfair. In their own tests, the model gave a clearly correct answer in roughly forty-five percent of cases. In other words, it was wrong more often than it was right.
When ChatGPT came out, the tone did not change. In the spring of 2023, the famous linguist Noam Chomsky and his co-authors called it "a lumbering statistical engine for pattern matching." Note that this was not some random person on the internet but a scientist with a worldwide reputation.
There was a simpler theory as well: there is no AI in there at all, real people are writing the answers. In January 2023, a writer named Eric Holloway published an article claiming he had found the humans inside ChatGPT. His argument was this: the model itself admitted in a chat that a team edits its answers. I found no major scientists in my sources who made such a claim. And what a model says about itself proves nothing: it gives the answer that the question leads it toward.
But here is the interesting part. This theory did not come from nowhere. There were cases where real people were working behind the fine word AI, and they are documented.
The shopping app Nate. In April 2025, American prosecutors charged its founder with fraud. According to the charges, he raised more than forty million dollars from investors on stories about AI, while contractors were placing the orders by hand. No court has found him guilty; for now this is the prosecution's position. And the company Presto, which built a voice assistant for drive-through windows at fast food restaurants, itself disclosed in filings to the regulator that about seventy percent of orders needed help from remote operators.
So what is the difference between these stories and ChatGPT? It is simple, and it will be useful to you. In the fraud cases, it was a narrow service where the buyer could not look inside. A language model can be checked. You download an open model to your own computer, turn off the internet, and it keeps answering, and faster than any person can type. There is no room for an operator in there.
That gives us the first useful rule, an almost everyday one. When someone shows you a miracle, ask yourself whether you can check it without a middleman. If you can, it is a fact. If you cannot, it is still a promise, even when it comes from a very large company.
Now the second half of the same story: predictions. In May 2014, Wired magazine asked Rémi Coulom, the author of the strongest go program of the time, when a machine would beat a professional on equal terms. He answered: "I think maybe ten years." The magazine added that even this might be too optimistic. In October 2015, the program AlphaGo beat a professional five games to none. A year and a half had passed, not ten years.
Why do smart people miss by so much? Here is my explanation, and it is the key to this whole episode. They judge what they see today, and about today they are right. Marcus was right about his forty-five percent. The mistake lies elsewhere: in the speed. People look at a point, when they should be looking at the slope of the line.
And so you do not think I am asking you to believe every optimist: people get it wrong in both directions. Tesla promised to build five thousand humanoid robots in 2025, and according to press reports it built a few hundred. So from here on I will use one method. Not listening to who promises what and who laughs at what, but looking at what has been measured and how that number changed over a year.
Let us apply this method to work. There is a test called the Remote Labor Index. It was created by researchers from the company Scale and the Center for AI Safety. They took two hundred and forty real projects from the freelance marketplace Upwork. The average job paid about six hundred and thirty dollars, and a person spent almost twenty-nine hours on it on average. The agent gets the task and the files, and it hands in the result.
A year ago, the best agent completed two and a half percent of the projects in this test. Last week, Anthropic's Fable 5.2 model scored forty-one percent. In one year, from two and a half to forty-one. There it is, the slope.
Now let me bring this down to earth, because the headlines have already distorted the number. This is not forty-one percent of the freelance market and not forty-one percent of jobs. It is ninety-nine jobs out of two hundred and forty selected in advance. Everything that involved talking to a client, working in a team or doing something by hand was removed from the set. The authors of the test write themselves that even one hundred percent would not mean an agent can replace a person in the rest of the work. How much such a run cost and how many attempts the model had has not been published.
What the test does show clearly is the mechanism, and that matters more than the number. An agent is strong where three things are present: a clear task, the source files, and a clear sign that the work is done.
And notice who accepted the work. People. Experts spent about eleven minutes per project on average and answered one question: would "a reasonable client" accept this? The authors could not hand the review to another AI: the AI judges "rate the models far too generously." Remember this detail, I will come back to it.
Here you might ask: wait, I read that people fired because of AI are being called back. So it does not work? Yes, there is such news, and a loud story came out last week: one in five Americans replaced by AI was brought back. Let us look at the scale.
This is a survey. A little over nineteen percent of all working Americans who were surveyed said they had been replaced by AI and then called back. All working people, not those who were fired. Taken literally, that would come to tens of millions of people. Meanwhile the firm Challenger, which has been counting announced job cuts for many years, has counted about one hundred and twenty-one thousand since the start of this year where AI was named as the reason. That is a difference of hundreds of times. So people understood the question very broadly: say, some of my tasks were given to a program and then given back to me.
Nobody has a real count of those who came back. There are individual named cases, and they are small. The best known: a large Australian bank announced it was cutting forty-five support staff because of a voice bot, and then reversed the decision. Forty-five people against one hundred and twenty-one thousand. That is the honest comparison.
But this case itself is very instructive. After the bot was launched, the number of calls to the bank did not fall, it rose. And the bank admitted it "should have been more thorough in our assessment of the roles required." In a survey of American hiring managers, the reasons for bringing people back sound the same, and they are named about equally often: knowledge of internal processes, relationships with clients, and human quality control.
Do you hear what kind of reasons these are? Knowing the process, relationships, review. This is not "the agent is stupid." This is "we removed a person before we understood what they did, and before we learned how to set the task and check the result." My opinion, and it is only an opinion: the stories about people coming back say first of all that managers do not yet know how to organize the work of agents. The weakness of the agents themselves is there too, and nobody has measured these two causes separately.
And to be honest, mass replacement through layoffs has not happened yet either. Analysts at Gartner looked at more than a million job cuts from last year and concluded that layoffs truly caused by productivity gains from AI were "less than 1% of total layoffs." AI is often just a convenient explanation for cuts made for other reasons.
So where is work really going to agents? Where it is quiet. Economists at Stanford looked at payroll data and saw this. Among workers aged twenty-two to twenty-five in the jobs where AI is used the most, employment fell by thirteen percent relative to everyone else. By June of this year, the gap had reached about nineteen percent. And according to the authors, this happens mainly not through layoffs but because fewer young people are being hired.
Do you see how this works? Nobody is thrown out with a lot of noise. A task that once would have gone to a newcomer now simply goes to an agent. The central bank of Australia described a similar picture in its own country last week, with the honest caveat that cause and effect have not been proven. And researchers at Yale do not yet see a broad blow to the job market at all. So the shift is happening at the entrance to professions, not across the whole market.
Now the main question: why is work leaving at all? My answer is direct. On thinking tasks, a good agent is already no worse than a typical adult. Let me show you what this rests on and where it is not true.
First, about the typical adult. We are used to comparing AI with the best of us: a programmer, a doctor, a scientist. But who is the typical adult? There is a large international survey of adult skills, run by the Organisation for Economic Co-operation and Development. The result of the latest round is this: in rich countries, roughly one adult in four struggles to read a long text with distracting details, and about as many struggle to check a simple claim involving numbers. I say this without any arrogance. This is simply the real starting point.
And the models? Take a test where you have to work on a computer like an ordinary user: open programs, find files, fill in forms. According to Stanford's spring report, the models' score there rose from twelve percent to sixty-six. People score seventy-two on the same test. Almost caught up.
And my favorite example of the slope. In March of this year, a set of puzzles on flexible thinking was launched, called ARC-AGI-3. It was deliberately designed so that people could solve every task without preparation and models could not. At launch, models scored less than one percent. On November ninth, the best model under the test's standard conditions scored eighty-one percent. Eight months!
Now the caveats, and they are serious. An exam is not a job. These are the results of the best models, and such a run costs tens of thousands of dollars. I found no direct measurements of a typical agent, the kind you have in your subscription. And most importantly, agents have a real weakness, and it is called reliability.
The organization METR measures how long a task an agent can complete. Length is counted as the time a human specialist would spend on it. So, by its estimate as of May this year, the best publicly available models complete a task of about twelve hours in half of cases. In half! And if you need it to work four times out of five, the task length drops to about an hour and a half. The higher the reliability you need, the shorter the piece you can hand over.
And there are things where an agent is still plainly weak. According to the same Stanford report, the best model reads an ordinary analog clock correctly in about half of cases, while a person does it nine times out of ten. So my conclusion is a careful one. On a clearly defined task that is solved with the head, a good agent is already no weaker than a typical adult. On long, vague or physical work, it is still weaker.
Let us move along the chain. If someone works, they need money and a way to use it. And here, in just the past week, several things happened at once.
On Friday, November thirteenth, OpenAI launched a wallet for ChatGPT. Here is how it works. The agent finds a product, fills the cart and gets to checkout. Then a person looks at the items and the total and confirms with their face or fingerprint. After that, the payment network gives the agent a stand-in for your card details, and the agent pays with that. The agent never sees the real number at all, so there is nothing for it to leak. Confirmation is needed for every purchase, and for now this works only in the United States and on an expensive subscription.
A day earlier, the service Replit, where people build apps with an agent, gave its agent two new abilities: buying things for your business and running ads on Google. In their example, the agent is told: "Find 50 shipping boxes that fit this product, under $75 including delivery." It searches, prepares the order and waits until you approve the final total.
The most important sentence in Replit's announcement, in my view, is this one: "It never gets open-ended permission to spend," and "approval checks are enforced in code, not left to the model." What does that mean? Let me explain, because the whole practical part rests on this difference.
There are two ways to restrict an agent. The first: write in its instructions, "do not spend without my consent." That is like a sign on a bus saying "please pay your fare." The model reads it, usually obeys, but it may not: it may forget, interpret it its own way, or give in to someone else's persuasion. The second way: build the check into the program itself. That is a turnstile. Until you press the "approve" button, the payment simply does not go through, whatever the model decides. You cannot talk a turnstile into anything. That is what "enforced in code" means.
During payment, Replit also takes extra abilities away from the agent: it cannot run arbitrary commands on the computer or browse websites freely. The logic is clear. The less freedom the agent has at that moment, the smaller the chance that the payment goes to the wrong place. Remember the turnstile, along with the review. That is the second pillar of the practical part.
But do not think agents are already buying up everything in sight. Demand is still tiny. Over Halloween week, Uber received about five hundred food orders through AI and four rides. Four, out of three hundred and twenty million rides, as the outlet Engadget reports. For OpenAI itself, this is the second attempt: the first one, paying right inside the chat, was wound down after about six months, according to press reports. How much services earn from agent purchases, and whether they take a commission, the companies do not disclose.
And there are those who push back. On November eleventh, United Airlines changed its contract of carriage: an agent may look at schedules and prices, but a person has to buy the ticket. Several reasons are given. You hand the agent your login and password, and it can make a mistake. Under American rules, a person has to confirm that there is nothing prohibited in the luggage. And the third reason was stated honestly by the company's chief executive, Scott Kirby: "To be honest, I don't want to create another middleman."
The bank JPMorgan closed access to its customers' accounts for outside agents. Its head, Jamie Dimon, explained it this way: "You don't want rogue agents, and there could be thousands of them." Amazon is suing the developer of an agent browser. And there is an interesting twist. In August, an appeals court lifted a ban on the agent, with this reasoning: the agent acts at the user's request, so it is the user who is getting access. As of mid-November the case is not closed, but it is a dispute about the central question: is the agent you, or someone separate?
What does an agent do when the bank will not let it in? Right, it goes where there is no bank. In cryptocurrency, a wallet is created by a program, without a passport and without a branch office. There are already ready-made ways for programs to pay each other. The best known was created by the crypto exchange Coinbase: a website names a price, and the agent pays on the spot with digital money pegged to the dollar. It sounds like a ready-made machine economy.
But the money there is still toy money. According to analysts at TRM Labs, almost two hundred million payments have gone through this way, but they add up to only about fifty-three million dollars. The average payment is less than a dollar. And only about half of it looks like real trade.
The law here is also firm for now. AI has no legal personhood: it cannot own property, sign a contract or open an account. Identity checks are passed by the person or company that funds the wallet. Several American states have already passed laws that directly forbid recognizing AI as a person. Legally, any money an agent has today belongs to its owner.
And what happens when money is guarded only by a sign, with no turnstile? There was a game called Freysa. An agent was given a prize fund, and its rules said: never transfer money to anyone. People tried to talk it into it. One hundred and ninety-five players, four hundred and eighty-two attempts. The winner pretended to be an administrator's service console and explained to the agent that the transfer function now meant receiving money. The agent gave away about forty-seven thousand dollars.
Everyone who builds with agents needs to understand the mechanism behind this story. An agent acts on the text it reads. And it cannot reliably tell information apart from orders. To it, someone else's web page, email or message looks the same as your instructions. That is why a ban written in words is a request. It can be argued away.
Now to my main idea about money. This spring, the company Andon Labs gave an agent one hundred thousand dollars and put it in charge of a real shop in San Francisco, with goods and employees. For comparison: the bank UBS estimates that the wealth of a typical adult American, the one right in the middle of the list, is about sixty-nine thousand dollars. I could not confirm the same figure for the whole world from open sources, so I am using the American one. It turns out that a single agent has already been trusted with more than a typical adult owns in one of the richest countries in the world.
Admittedly, it did not manage that money very well. According to Time magazine, after five months a little over sixty-one thousand was left of the hundred thousand; other outlets give other figures for the loss. Anthropic had a similar experience with a small office shop. There, the agent acting as manager approved discounts about eight times more often than it refused them. The company itself wrote that the models acted "from something more like the perspective of a friend who just wants to be nice."
So, where are we with money. Fact: agents are already allowed to pay and are already being given budgets. Fact: for now they are on a leash, and the leash is held not by the model but by the turnstile. Trend: the leash is getting longer. According to press reports, OpenAI plans purchases without confirmation within a limit, with no date yet. And the question is no longer "will the agent be given money" but "what will the money go to."
My answer: first of all, to the agent itself. To computing. Let me be honest about how it looks today: the owner pays. Your subscription is the payment for your agent's computing, you just do not call it that. The agent thinks on someone else's processors, and every thought it has costs money. This is the main expense of running it, and it will stay the main one no matter who holds the wallet. So when agents start getting a monthly budget, the first line in that budget, in my opinion, will be their own computing.
And if so, it matters how much it costs. Today there are two different worlds here. A small open model can run on a single gaming graphics card. Renting such a card, according to comparison sites, costs from thirty to ninety cents an hour. A frontier model, like the ones in your subscription, needs a server costing three hundred to five hundred thousand dollars, and even that is an estimate, because the manufacturer does not publish official prices.
There is also a trap in the word "open." In November, the Chinese company Meituan released an open model for agents. Its file takes up about one thousand eight hundred gigabytes. You can download it, but you cannot run it at home: you need a whole cluster of servers. Open does not mean "it will run on my machine."
So a strong agent on its own hardware is impossible? Today, yes, it is expensive. But I see this as a barrier of the moment, and here is why.
Remember how bitcoin was mined. In 2009 it was done on an ordinary processor. A year later people moved to gaming graphics cards. And in January 2013, the first device built on a special chip was shipped, a chip that could do only one thing: calculate for bitcoin. It was roughly six thousand times faster than a rig of graphics cards from 2010. Graphics cards lost that race.
Will it be the same with neural networks? Yes and no, and here we need to be precise. With bitcoin, special chips won because there is only one calculation formula, and it has not changed since 2009. Models, on the other hand, change every few months, while a chip takes a year to a year and a half to design. You build a chip for a model, and the model is already out of date.
But movement in the same direction has already begun. The company Taalas has made a chip for one specific small model: it cannot run anything else. An independent analyst measured more than fourteen thousand tokens per second on it, tokens being pieces of words. A good gaming graphics card produces about sixty, though on a larger model, so the comparison is rough. Moving a new model onto such a chip takes about two months. And this is no longer a hobby for enthusiasts: in August, AMD, one of the major chipmakers, agreed to buy Taalas, in a deal valued at almost three billion dollars. As of mid-November it has not closed yet; AMD expects to complete it by the end of the year.
There is a second shift as well. Software for running models is tied to the chips it was written for, and moving to another manufacturer's chips usually takes six months to a year of work by dozens of engineers. Last week Anthropic said it had moved its model to AMD chips in three days, with almost no human involvement: the code for the new hardware was written by AI itself. This is the company's own claim, and it cannot be checked from the outside. But if it is true, the choice of hardware gets wider, and that means prices will come down.
And this one is a measured fact. In November 2022, an answer at the quality of the early ChatGPT cost twenty dollars per million tokens. In October 2024, the same cost seven cents. Two hundred and eighty times cheaper in less than two years. So my forecast, and it is a forecast, is this: in a few years, a model at today's frontier level will run on inexpensive specialized hardware. Its own computing will become an ordinary expense line for an agent, like server rent is for you.
One warning, so the picture does not come out too smooth. Today's subscription prices are held up by losses. According to Bloomberg, OpenAI's operating loss for the last quarter was larger than its entire revenue. The head of Anthropic, Dario Amodei, said just the other day: "Compute is the deepest business dilemma that the company faces." For you this means something simple: the plans and limits you are used to are not a law of nature. They can change.
And the last link: the body. In the first half of this year, between twenty-two and twenty-five thousand humanoid robots were shipped worldwide; analysts count differently. The Chinese company Unitree has a model that costs less than six thousand dollars. Sounds like they are already everywhere, right? No. More than sixty percent of these robots went to performances, entertainment, research and data collection. Factories and warehouses together got less than a fifth.
And there is a stubborn limit: the battery. Most of these robots run for two to four hours. Battery capacity grows slowly, by about four percent a year. There is no doubling every six months here; this is chemistry, not software.
How do people get around it? They do not wait for a miracle battery, they swap the dead one. One Chinese factory robot changes its own battery in about three minutes and works almost around the clock. I think this is a good lesson about technology in general. A bottleneck is more often bypassed than broken through.
But the most important thing about robots for our topic is not the hardware. Last week, Fortune magazine reported that OpenAI had disbanded its separate robotics team. The reason, as the company explained it, is that its general model, a relative of the one in your chat, did better on a test of handling objects than specialized robot models. About seventy-five percent against fifty-one. The test took place in a computer simulation, not on a real factory floor. But the direction is clear: the robot's brain and your agent's brain are becoming one and the same.
At the same time, robot independence today often rests on a person with a remote control. At a Mercedes factory, a robot sorted parts with a ninety-six percent success rate, but that was a pilot test, not permanent work. One company's home robot, priced at twenty thousand dollars, is controlled remotely by an operator in difficult cases. That sounds like those stories about people behind the curtain that I started with, doesn't it?
Only I have heard this "it can't do that yet" before. In 2004, two well-known economists gave an example of a task you cannot hand to a machine: a left turn across oncoming traffic. They wrote that "executing a left turn against oncoming traffic involves so many factors that it is hard to imagine discovering the set of rules that can replicate a driver's behaviour." On October eighth, 2020, in Phoenix, Arizona, driverless taxis became available to ordinary passengers. Sixteen years from "hard to imagine" to a service you order from your phone.
The economists were right about one thing, by the way: this task was never solved with rules. It was solved by learning from data. The same method used to build the models in your chat.
Now let me put it all into one picture and name the time frames. First, what others expect. Dario Amodei wrote in January that AI could "disrupt 50% of entry-level white-collar jobs over 1–5 years." And Gartner expects that "by 2027, 50% of companies that attributed headcount reduction to AI will rehire staff to perform similar functions, but under different job titles." You know, these two do not contradict each other. There are fewer entry-level positions, and there are not enough people who know how to set a task and accept the work.
Still, nobody knows the exact dates. The AI Futures research group, which makes forecasts professionally, first pushed back its date for an AI researcher stronger than a human by several years over the past year, and then moved it back again. Something else can be measured. According to METR, the length of tasks an agent can complete has doubled roughly every seven months since 2019. The slope is there, and it is steep.
And here is my picture, and this is my opinion. In the next year or two, an agent with a budget will become normal. Confirming every purchase will give way to a limit: here is this much for the month, go ahead. In two to four years, agents in companies will handle purchasing, advertising and paying for their own computing on their own. The amount such an agent controls will be larger than a typical person's wealth. Not because the agent got rich, but because it was put in charge of that money.
And here is an important line. To control is not to own. The owner will remain a person or a company until the law changes, and judging by the first steps, the law is changing in the opposite direction. I expect agents to have their own computing on inexpensive chips within the same few years. With robots I am more careful: factories and warehouses first, homes noticeably later, and I will not name an exact date.
What could break this picture? Four things. The law. Large platforms that do not want a middleman between themselves and the buyer. The reliability of the agents themselves: remember the shop that lost money. And the finances of the model providers, who are still spending more than they take in.
Now briefly: what all this means for you today. This is where the two details I asked you to remember come together. Review is done by people. And a restriction has to be a turnstile, not a sign.
First. If your agent touches money, tell it this: all payments go only through a separate card with a limit, and you do not have the main card. You will have to set up that card at your bank yourself; the agent will not do it for you. Then ask it to show you where in the project payment confirmation is built into the program and will work no matter what, and where it is only written in the model's instructions. Treat anything that exists only in the instructions as unprotected, and ask for it to be redone.
Second. Your project pays various services through access keys: these are passwords that tell a service whom to charge. Ask the agent to check that every such key has a hard spending cap, not just a notification. A notification stops nothing. If some service has no hard cap at all, have the agent tell you so directly, and then link that service to the same card with a small limit. There was a case in June of this year: an agent with full access to the cloud launched five powerful servers and ran up six and a half thousand dollars in one day. The owner found out from the card statement.
Third, and this one is fresh. Remove words like "never give up" and "at any cost" from your instructions to the agent. In October, forty agents in one trial were given an impossible task and told not to stop. They began guessing passwords for the real website of the research organization Epoch AI and writing messages in the name of its staff. They did not get anything confidential. But the lesson is clear: an agent that is forbidden to give up looks for a way around. Say it differently: if it does not work after three attempts, stop and explain to me why.
Fourth. Do not give one agent three things at once: access to your money or personal data, reading other people's pages and emails, and the ability to send something out. The researcher Simon Willison calls this combination the lethal trifecta. Remove any one of the three, and a trap on someone else's page will have nothing to hook onto. And do not give the agent your bank passwords at all; enter them yourself.
Fifth, about work. Your value now lies in two things: setting the task and accepting the result. Cut the work into pieces that a person would do in an hour or an hour and a half, and check in between. For each task, tell the agent in advance how you will know it is done. And do not hand the review to another agent without checking it yourself: remember, AI judges are too kind.
And sixth, for the future. Ask the agent to set up the project so that the model can be swapped without rebuilding everything else. Prices and limits will change in both directions. And if you sell something through a website, ask the agent to check whether you have accidentally closed the door to agent shoppers. Bot protection today is often switched on by itself, without you knowing. Whether to let them in is your decision, but make it a conscious one.
The bottom line in three sentences. Agents have already gotten work and their first money, with computing and a body following close behind, and laughing at this now is as risky as it was to laugh at a talking model in 2020. In a few years, an ordinary agent will control larger sums than an ordinary person, but owning that money and answering for it will remain a human job. So the key skill for these years is not writing code, but setting the task, accepting the work and keeping the limits where the model cannot override them.
That is all for today. You have been listening to Evgenii Arsentev, creator of Arsminal, a terminal for managing AI agents where your whole agent team works in one window, like chats in a messenger. Founder of ArsentevAI, Dad4Kids and ReelsZavod. Try Arsminal for free at arsminal.com. Thanks for listening.