← AI Agents & Vibe Coding
Episode 13··30:47

Agent over the wire: Claude Code in phones, cars and TVs, and who is making money on it

You have Claude Code or Codex, and on social media you see people plugging an agent into a phone, a car or a TV with a cable. It is hard to tell where the toy ends and a new business begins, and whether all of this will be shut down in a month. From this episode you will understand how an agent actually controls someone else's device and why this started working right now. You will learn how startups make money on cloud phones, what it costs, and where the line runs between this and the gray zone of account farms. I break down what car makers, Google and Xiaomi are already closing off, and what, by all signs, they will leave alone. And I give you rules to start with on your own: one device on your desk, read only at first, confirmation before any irreversible action.

Transcript

Tap a paragraph to listen from there.

Good evening, startup founders and vibe coders! Today I am going to tell you how an agent like Claude Code or Codex stops living only inside your computer and travels down a cable into someone else's device: a phone, a car, a TV, a scanner, a robot. And most importantly, I will tell you who is already building a business on this, how it works inside, whether it will get shut down, and where it is all heading.

Let me start with a picture almost everyone remembers. Star Wars. A small round droid rolls up to a wall, extends a probe from its body, plugs it into a socket, turns something inside, and the gate opens. Nobody in the cinema asks how exactly it came to an agreement with that gate. It simply connected, and the foreign machinery obeyed. For many years this was a fairy tale. But the last few months on social media look as if someone quietly put that fairy tale on sale. People take a laptop with an agent on it, plug a cable into a phone, a car, a TV, and write in plain words: figure out what this device is and make it do what I need. And the agent figures it out.

Let me be honest right away, so I do not disappoint you later. In the movies, the droid magically understands any door. In real life, things are more modest and more interesting. The agent has no special power over hardware. It does exactly what a capable engineer would do, only very fast. It reads documentation and forums, tries commands, writes small translator programs, looks at what the device answered, makes mistakes, corrects them. The pattern in every story I am going to tell you is the same. The agent has a terminal. In the terminal there is a program that can talk to the device in its own language. And the agent finds that program, writes it, or finishes it. That is the whole droid probe. Remember this, because everything that follows, where the money is and where the danger is, depends on understanding it.

The loudest story of the week is about a car, and I will get to it. But I will start with what matters more for you, which is business. People shoot videos about cars, but right now the money is being made on phones.

Think about what a phone means for a business. It is a door to almost everything: social networks, messengers, banking apps, thousands of services that have no proper entrance for programs, only an app. If an agent can hold a phone in its hands, scroll, tap, type and read the screen, it gains access to a world that used to let in only a human with a finger.

How does an agent hold a phone? Every Android device has a service entrance for developers. Officially it is called the Android Debug Bridge, and from here on I will simply call it the debug bridge. You turn it on in the settings: first developer mode, then debugging over a cable or over the network. After that, a computer can install and disable apps, take screenshots, tap at a given point, enter text and, most importantly, read the tree of elements on the screen. That means not just a picture, but a list: here is a button labeled "Publish", here is a text field, here is a list of videos. For the agent, this is like giving a blind person a description of the room. It does not guess from pixels, it knows where everything is. This works the same way on phones and on Android TVs, which is important and will come in handy later.

The iPhone is harder, and here you can see how inventive people are. Apple does not provide such a service entrance. So people use a workaround: on a Mac they open the iPhone screen mirroring window, the agent looks at that picture, recognizes the text with Apple's built-in tools, and clicks by coordinates. In other words, the agent connects to the iPhone not with a cable, but with its eyes.

The first startup worth talking about is called Phone Harness. Its founder, Shawn Pana, used to run growth at Browser Use, a company that teaches agents to work in a browser. The logic is clear: once you have taught an agent the browser, the next door is the phone. The project launched in early November, and it is part of this year's fall batch of the Y Combinator accelerator. Its core is open, it sits on GitHub under a free license, and today it has about five thousand one hundred stars, that is, bookmarks from developers, which is a lot for a project like this. The promise is simple: any agent, Claude Code, Codex or another one, can "scroll, tap, type, and read your phone", without jailbreaking the system and without installing apps on the phone itself. Android through the debug bridge and the screen tree, the iPhone through that same mirroring window on a Mac.

And here is what I like about this project: it honestly says what it cannot do. If you unlock the iPhone, the mirroring stops. An Android phone with a PIN code needs a human to unlock it. The agent will not understand icons without labels, that needs a model with vision. There is no multi-finger touch, no camera, no Face ID, and protected video shows up as a black rectangle. So this is not the droid from the movie. It is a pair of very nimble hands with a few fingers tied together.

How does the company make money? The core is free, and the paid part is the cloud: you do not need to keep a stack of phones on your desk, you rent them remotely. Cloud Androids by the minute, real iPhones by subscription. The prices on their official page are as follows. A cloud Android costs five cents a minute. A real iPhone in the cloud is eighty dollars a month, cheaper if you pay for a year. And if it needs an eSIM with a real phone number, it is one hundred dollars. And notice how the cloud is advertised: "stealth, real numbers, 2FA". Those are exactly the words you need where platforms hunt for bots. The developers themselves say nothing in the project description about social network rules or bans. The sources also say nothing about funding, apart from taking part in Y Combinator.

The second player is older and more interesting for us Europeans. A German startup that used to be called Droidrun and is now called Mobilerun. In the summer of last year it raised two point one million euros at the very earliest stage, the lead investor was the fund Merantix Capital, plus one more fund and private investors. At the time of the round they had a very fast start: more than nine hundred developers in the first twenty-four hours and more than three thousand three hundred stars on GitHub in under three months. In essence they do the same thing: an open framework through which an agent controls Android and the iPhone in plain words. And what they sell is the cloud: cloud Androids, real iPhones, numbers for text message verification, email inboxes, and a ready-made connector for Claude and the Cursor editor. Their prices are public: a free plan, then twenty, ninety-nine and two hundred ninety-nine dollars a month, and for corporate fleets by agreement.

The third branch is quieter, but for many of you it is more useful. The startup Limrun, also a Y Combinator alum, from this year's spring batch. It gives agents cloud simulators of the iPhone and Android, and a remote Apple development environment. This is not about farms. This is about an agent that is writing a mobile app for you being able to launch it on its own, poke around, find where a button has drifted, and fix it. There is a free trial plan, and paid plans start at sixteen dollars a month.

And now the most uncomfortable but honest part of the market. There is a whole layer of Asian companies that have long been selling cloud phones for social media. The information about them comes mostly from their own press releases and advertising, so take it with a grain of salt. GeeLark has released an official connector through which an agent, Claude Code, Codex or Cursor, creates a cloud phone by itself and, for example, scrolls TikTok. It works on plans with programmatic access. VMOS Cloud advertises a connection to Claude Code with a scenario for warming up TikTok accounts, that is, training a new account to behave like a real person, and promises Reels, Shorts and WhatsApp next. DuoPlus has an agent built right into the service, taps and swipes by words. And in open source there are a dozen free bridges between the Android debug entrance and an agent. Among them, for example, is a project for warming up TikTok accounts, which its author honestly calls "a 3-hours hack project mostly written by an LLMs".

What follows from all this? The phone farm market already sells itself through compatibility with Claude Code. A farm used to be a wall of phones and a script that pressed buttons at fixed coordinates. The script broke with any app update. Now, instead of a script, there is an agent that reads the screen and decides by itself where to tap. That is cheaper to maintain and more flexible. And that is exactly why there is so much money and so much gray zone here at the same time.

Look at all these companies together, and you will see the same business pattern. First, a free open bridge: a project on GitHub, thousands of stars, compatibility with Claude Code and Codex. That is the marketing, they spend almost nothing on advertising. A developer finds the project, tries it on their own phone, likes it, and when they need ten, a hundred, a thousand devices, they go to the paid cloud, by the minute or by subscription. That is the path of Phone Harness, Mobilerun, the car startup I will talk about later, and the Asian cloud farms. They get customers not through advertising, but because the developer's agent already knows how to work with them.

And now the car I promised. Briefly, because this is not business, but the most vivid example of how an agent travels down a cable into hardware.

There is a man named Muhammad Zulali. He has a nineteen ninety-four Jeep Cherokee, a car that is thirty-two years old. The car was failing the California emissions test: under a certain load, at around twenty-five miles per hour, it was putting out too many nitrogen oxides. The trouble with old cars is that they do not have the standard diagnostic port that came later. To get into the brain of a Jeep like this, you have to speak Chrysler's own language. Zulali took a dirt-cheap Arduino board, about three dollars by various estimates. The whole kit came to fifteen dollars at most. The board went into the old diagnostic port. Claude explained that Chrysler has its own protocol and that a translator was needed between the car and the laptop, helped with the wiring, and started pulling live data straight onto a MacBook.

Then something curious came out. The California firmware of the engine control unit cut fuel under loads similar to this test, and that was getting in the way of passing. You cannot rewrite that firmware the usual way. So the second agent, Codex, wrote a small program right into the unit's memory. The car, according to its owner, passed the test. As the outlets report it, Zulali put it roughly like this: "Claude did most of the diagnostic work and Codex did pretty much all of the memory work." The post titled "Claude reverse engineered my 32-year-old Jeep. Then it started controlling it." got about one point eight million views, and car publications picked up the story.

And now the part that usually disappears in the viral retellings. Claude's first hypothesis was wrong: the agent decided the engine was getting about twenty-two percent too much fuel, but the culprit turned out to be a tool taken from the internet that was giving bad data. The agent trusted a faulty instrument, just as a person would have. The second point is more worrying. According to one observer's retelling, Claude itself suggested turning on the fuel pump and turned it on, and the pump ran for about ten seconds until the owner stopped it. The Drive puts it more gently: the model "asked to run the fuel pump". Either way, a program on a laptop reached a real pump in a real car.

And a little more sobriety. Readers of one car publication, people who know these Jeeps well, write that the root cause was never found: it could be the pump, the filter, the injectors, an air leak. So, in their words, this is a band-aid, not a cure, and some of the readings in the video look doubtful to them. These are opinions in the comments, not the editors' conclusions, but it is a useful reminder: not every number on the agent's screen came from the car itself. And one last thing. In America, a firmware change that affects emissions can be considered tampering with the emissions control system. None of the publications I read examines whether this story was legal. So this is an open question, not a recipe.

There is also a calmer example. A British man from the channel Yotty Labs bought a wireless diagnostic adapter on Temu for two pounds thirty-nine pence, plugged it into his father's car, and asked Claude Code to build a dashboard. The agent figured out on its own how the adapter talks to the laptop, wrote and tested Python scripts, and brought up a dashboard in the browser with temperature, revs and error codes. The author says: "I've been programming for over 20 years, and it's just done that in 5 minutes." Doing it by hand, by the author's estimate, would have taken almost a day. And very honestly adds: "We haven't actually hacked the car in the traditional sense." That is true. The standard port gives out standard data. The agent saved time, but it did not open a secret door.

From this part, remember one line from a man named Kai Wang, who took this wave apart: "A bad commit, you revert. A car you just broke, you don't." We will come back to it. But for now, let me tell you how people are already trying to build businesses around cars.

The first example is from Denmark. The company CSS Electronics makes loggers that record the car's internal network, the one its control units use to talk to each other. They released free skills for Claude Code that help decode the manufacturer's closed signals, that is, understand which message on this network means, say, the pedal position. But these skills are built around their own device, although they can be adapted to other adapters too. This is the second business pattern, subtler than the first. They still sell hardware, but the skill for the agent has become a sales channel. The buyer comes not because they saw an ad for the logger, but because their agent knows how to get along with that logger.

And the car startup I promised. The project is called car-use, it is made by an organization called October, and their motto is simple: "Let your agent use your car." It works with Claude Code, Codex and a couple of other agents, through cheap diagnostic adapters for about fifteen dollars. But the most interesting part is their rules. By default, read only. Writing, for example clearing error codes, only after an explicit "yes" from the owner. Brakes, steering, the odometer, disabling emissions systems, keys and driving itself are always forbidden. Every action is written to a log. There is a virtual car with seventeen breakdown scenarios, so the agent practices on it rather than on your vehicle. The paid part, a cloud with permanent memory for each car, is still in early access, and it mentions trucks, cranes and generators. But let us bring it down to earth right away. Today their project has seven stars. Prices, founders and funding are not listed. This is a very early project with a good idea, not a company with customers.

About TVs, very briefly, this is a household story, but it showed the scale. The wave was in October. A man sent Claude Code a link to his TV model and asked whether it was possible to "hack this". The agent removed the ads, the junk apps, the trackers and the lag, and made the screensaver a screen with revenue, a calendar, the weather and Stripe payment notifications. Another post, about parents' TV that showed ads before every app, got almost nine million views: the agent found the TV on the home network, removed twenty-five preinstalled Xiaomi apps and made its own home screen. This works through the same debug bridge, so only on Android TVs. Samsung, LG and Roku cannot be handled this way.

Besides phones and cars, there are a dozen quieter stories, and they show how widely this is spreading. A well-known Linux software developer, Christian Schaller, had a film scanner that stopped working after the manufacturer changed the chip inside. Claude Code compared what the Windows driver and the Linux driver were sending to the scanner, and within a few hours there was a first scan, and within a few days an almost complete driver. The same approach fixed the webcam in a laptop, with patches ready in a day. The owner of a Bluetti home power station decoded its protocol and connected it to a smart home without the manufacturer's cloud. Engineer Lucas Gerads connected an agent to an oscilloscope, and the agent checks a circuit simulation against the real board. The lesson is worth a lot: do not let the agent guess what is connected where, and make sure it does not use old measurements.

Now let us talk about why all this started working right now. After all, the Android debug bridge has existed for more than ten years, the diagnostic port in cars has been mandatory in America since nineteen ninety-six, and adapters for it cost from two to fifteen dollars. The hardware has been around for a long time. What was missing was someone to sit by the cable, read the forums, work out the protocols, write the translator and patiently keep trying. That used to be an engineer, whose hour was expensive, and so home electronics stayed closed not because they could not be opened, but because opening them did not pay. The agent brought the cost of that patience down to zero. And suddenly it turned out that a huge amount of the technology around us is slightly open: the service entrance is there, it is just that nobody wanted to spend time on it.

There is a second half to the explanation. A single way to connect tools to an agent has appeared. It is the Model Context Protocol, invented by Anthropic, and the easiest way to picture it is as a power socket. Someone writes a set of actions for a device once: read error codes, take a screenshot, press a button. And any agent that knows this socket can use those actions. This is more convenient and safer than letting an agent into a bare terminal, because you can allow only some of the actions. For example, reading is allowed, but erasing is not.

Now the question you have probably already asked yourself: will all this be shut down? The picture is different for different doors, and I will tell you what is visible now, without prophecies.

Cars have been closing up for a long time, and it has nothing to do with neural networks. At Chrysler, Dodge, Jeep and related brands, since roughly the twenty seventeen and twenty eighteen models, there is a security gateway between the diagnostic port and the car's internal network, and only authorized tools are allowed to write anything. That is according to industry and advocacy publications. The general trend: data is moving into the car's encrypted cellular link with the manufacturer, and the port under the steering wheel is losing its importance. On the other side is the fight for the right to repair. A federal law on this in America, judging by the latest reports, has still not been passed, and the state of Massachusetts has a law that requires standardized access. So the story with a thirty-year-old Jeep is possible precisely because the car is old. With a new one, it will not be so easy.

Android is more interesting. Google really is tightening the screws, but not where you think. Since the end of September, a new rule has been in force in Brazil, Indonesia, Singapore and Thailand. If a developer has not passed identity verification, their apps are blocked when installed from participating stores. Worldwide, this is promised for next year. For experienced users there is a workaround with a reboot and a twenty-four-hour wait. But installing through the debug bridge remains an official route, with no waiting. And I did not find any plans in the sources to close the debug bridge or network debugging. The logic is clear: all Android development rests on this entrance, and closing it would mean shooting yourself in the foot.

Xiaomi, by the way, has put up its own barrier, and everyone building something on their phones should know about it. According to third-party guides, not Xiaomi's own documents, their firmware has, besides the usual cable debugging, a separate switch: debugging in the security settings. Without it, a computer can look, but cannot tap or enter text. To turn it on, you need to be signed in to a Xiaomi account with a SIM card inserted, and then sit through three warnings of ten seconds each. So the manufacturer already admits that controlling a phone over a cable is a separate level of trust. But the barrier does open.

TV manufacturers are silent so far. I found no statements from Google, Sony, TCL, Xiaomi or Samsung about this wave, and no reports that anyone has closed developer mode.

Social networks will fight, but not against the cable. They do not care how a button was pressed. TikTok's rules directly prohibit using automation to create and run accounts at scale and to inflate engagement. And you need to understand the scale of this fight: according to TikTok's own report for the first half of twenty twenty-four, they prevented more than thirty-six billion fake likes and more than seven hundred million fake accounts. I have not checked fresher numbers, but the order of magnitude is clear. When a cloud farm seller promises zero percent bans and complete invisibility, that is advertising without a method behind it. The platform catches not the cable, but the behavior: identical actions, identical timing, unnatural chains of events.

So my answer to the question "will they block it" goes like this. The cable will most likely be left alone, because developers depend on it. Blocking will happen at two other levels. The first level is writing to critical systems, as with cars, where reading is allowed but writing is only for the authorized. The second level is accounts and behavior, as with social networks. The droid will be allowed to plug in its probe, but the gates will open only for those who are supposed to go through.

Now the most important part. What can someone do with all this if they are not a programmer, but already have Claude Code or Codex and want to launch their own project? I will give you only what follows directly from these stories, and I will not repeat general advice.

First. If you have an idea that needs a phone, and the service has no proper programmatic entrance, you no longer need a programmer to test the idea. Take an old Android phone, turn on developer mode and cable debugging, and on a Xiaomi also turn on that second switch in the security settings. Connect it to your laptop and ask the agent to install one of the open bridges, for example the Phone Harness core or the Mobilerun framework. Then you write in words: open this app, find this screen, read me what it says. That is an evening of work, and you can already see whether it will work or not. Do not build a cloud and do not buy ten phones until one phone on your desk has done your task from start to finish.

Second. Build into your plan from the start the limitations the developers themselves honestly write about. A phone with a PIN code will need a human. The iPhone works only through the mirroring window on a Mac, and unlocking it breaks the connection. Icons without labels will need a model with vision, and that is more expensive and slower. If your idea depends on the camera, Face ID or multi-finger touch, this path is not for it. Better to learn that on the first evening than a month later.

Third. Do the math properly. The hardware here costs almost nothing: a car adapter is two to fifteen dollars, and you already have an old phone. A cloud Android, according to the official Phone Harness page, is about five cents a minute. That sounds cheap, but do the math: one phone running around the clock is more than ten thousand minutes a week, which is about five hundred dollars a week and more than two thousand dollars a month. A real cloud iPhone is about eighty dollars a month. And the main cost is something else entirely: your time spent on confirmations, and the price of a mistake. An agent that needs your "yes" for every important action needs someone to watch over it.

Fourth, and this is the most important. The safety rules you have seen in these stories, do not invent them anew, just copy them. Companies that know what they are doing have arrived at the same rules. The car startup allows only reading by default, and writing only after an explicit "yes". Industrial developers who connect agents to machine tools say the same thing: read only at first, writing goes through a human. A project for routers can roll back its settings by itself on a timer if something goes wrong. A project for 3D printers requires confirmation before printing and before changing the temperature. And the enthusiasts who cleaned up TVs came up with their own rule: "disable, don't uninstall", and "Everything is meant to be reversible." For them, removing system packages broke the home button and input switching. So tell your agent right in the first message: do not delete anything, do not erase anything, disable instead of deleting, and before any action that cannot be undone, ask me.

Why do I insist on this so much? Remember the pump in the Jeep that the agent turned on by itself. And there is one more user report, and I stress, it is a single unconfirmed one: an agent on a developer's phone decided on its own to clear an app's data for a clean test, and wiped about four gigabytes of unique models the person had been training. The agent is not evil. It is just very diligent about its task and does not always understand what is irreplaceable for you. "A bad commit, you revert. A car you just broke, you don't."

Fifth. Close the door behind you. Turn off network debugging on your phone or TV when you are done. If you leave it on, an open service entrance remains on your home or office network, and the device can be controlled through it. This is a direct warning from the same sources that described cleaning up TVs.

Sixth, about keys. The other day the cybersecurity company Darktrace showed something unpleasant. If an attacker has already taken over a computer and stolen the key to the agent, they can turn Claude Code on that computer into their own remote control. In their lab network, such an agent scouted the network, found a password, tried it against user accounts, and exfiltrated a one-gigabyte file. Important: they found no vulnerability in Claude Code itself, the hole was that the machine had already been hacked and the key stolen. The conclusion is this: an agent with access to your devices is a trusted employee with keys to every door. Treat it accordingly. Do not keep the agent's key where other people have access. Do not give an agent that is connected to phones with your accounts extra rights to everything else.

Seventh, about farms, since there is so much talk about them. If you are building a content factory on your own phones and publishing your own content to your own accounts, the agent really does take a huge chunk of routine off your hands. But keep in mind that TikTok's rules prohibit automation for creating and running accounts at scale and for inflating numbers, and that this fight is measured in billions. Sellers' promises of zero bans are worth nothing. If your business depends on the platform not noticing you, it depends on someone else's mistake. If it depends on the agent saving you hours of honest work, it depends on you.

Eighth, for those looking for a startup idea. Look again at the pattern that repeats across everyone. A free open bridge to some class of devices that works with Claude Code and Codex, and the money is in the cloud, in the device fleet, in memory, in the action log. Phones are already taken, and densely so. But cars are only just getting started, the most visible project has seven stars. For printers, routers, office equipment and lab equipment, there are no viral stories at all yet. And the second pattern, the Danish one: if you already have a product, hardware or a service, build a skill or a connector for agents for it. Your next customer may come not through advertising, but because their agent already knows how to work with you. For a small team, this is perhaps the most underrated sales channel right now.

And ninth, a very small one. If there is an Android TV hanging in your office, turn it into a business dashboard with sales, payments and a calendar. That is an evening of work for the agent, and a good way to learn, on a safe device, everything I have talked about: the service entrance, reading, careful writing, reversibility.

Now, about where all this is heading. Here I will be careful, because I did not find independent analysts on this exact topic, and forecasts come mostly from those who make money on it.

At the end of the summer, Anthropic released a preview version of what it calls a standard for hardware. In short, it is an attempt to give physical devices the same socket that already exists for software. The device describes itself, commands are reduced to simple reading and writing, and safety limits are written in plain human language: for example, do not heat above a certain temperature. It all started at a research institute. Among the partners are the pharmaceutical company Genentech, universities, Amazon's cloud, makers of lab robots and industrial robot arms, and among the early adopters are Hugging Face and Raspberry Pi. The results, and these are the company's own claims, are as follows. Connecting a device shrinks from weeks and months to hours and minutes. At one university, experiments ran about three times faster. And at a company that builds quantum computers, an AI controller restored the laser setup by itself in ninety-nine point three percent of cases. At the same time, the limitations are named directly. The standard does not work with devices that have no programming interface, the model's spatial reasoning results "still require expert oversight", the code is not open yet, access is by application, and the physical safety plan is still in progress.

And one more story from the same company, which is a good reality check. In the summer they ran an experiment with a robot dog. The model, working alone, completed a set of tasks to set up the robot: human teams without AI help spent three hundred sixty-one minutes on four shared tasks, with AI help one hundred eighty-one minutes, and the model alone nine minutes and thirty-five seconds. That is almost twenty times faster than the best human team. It sounds like the start of science fiction. But the fetch task itself, accurately pushing a ball, the model did not manage. In the first phase of this experiment, a year earlier, the robot once ran off and nearly crashed into the team. And it was still a human who plugged in the cable, a human who gave the first request, a human who approved the commands. Anthropic itself writes: "We are plausibly entering the early era of physical agentic AI." But it immediately cautions that "there may be substantial barriers" along the way.

If you put it all together, the future looks like this. Right now we are in the era of homemade probes: every enthusiast writes their own translator for their own device, and it is fun, cheap and a little dangerous. The next step is already visible: devices will start telling the agent themselves who they are, what they can do and where their limits are. This is needed first of all by labs and factories, where the agent is let in through a layer that physically prevents it from bypassing a machine's safety interlocks. Household devices will follow, or they will not, that depends on the manufacturers, and they are silent for now. And in parallel, the doors behind which there is money and safety will be closing: cars are already hiding writing behind gateways, and social networks catch behavior, not the cable.

What follows from this for you is a fairly down-to-earth thing. The window in which a small team with an agent can build a bridge to a whole class of devices and become the entry point for it is open right now, before the manufacturers do it themselves. And the advantage will go not to those who are first to connect an agent to everything in sight, but to those who are first to do it in a way that makes the device owner unafraid to say "yes".

In the movies, the droid plugged in its probe and the gates opened. In real life, the agent can already plug in its probe too, only first it spends a long time reading the manual for the gates, sometimes mixes up the buttons, and sometimes turns on the pump without asking. The gates do open, but for now a human has to keep a hand on the switch.

Let me sum up. Agents have learned to work with other people's devices not because they gained magical powers, but because for pennies they do an engineer's patient work: they read, write translators, try things and correct themselves. The main money right now is in phones, and all the players follow the same road: a free open bridge and a paid cloud, and what will get blocked is not the cable, but writing to critical systems and account behavior. If you want to build on this, start with one device on your desk, read-only mode and the rule "ask me before anything irreversible", and look for your idea where nobody has built a bridge to the devices yet.

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.

0:00−30:47 left