Transcription
Anthropic finished up shipping OpenClaw for Claude. What they shipped is dispatch and then a day later computer use. And what they give you together is exactly what people do with OpenClaw but without all the security risks. So you can text Claude from your phone. It takes over your desktop. It can open up apps. It can click its way through screens. It can navigate tools that have no API. And it delivers finished work while you are away at your phone wherever you want to be. Not a summary for you to review, not a draft to edit. The thing itself just gets sorted. And that distinction between work that lands on your desk and work that gets off of it is the whole game now. And I don't think we're talking about it enough.
So much of the time I see AI agent demos that are optimized to look good, not optimized to actually get work off your desk. Like, let's be honest. How many AI agents want to come to you now and say, "I have a briefing for you, a proactive briefing. It's a really important briefing. It's going to be before your meeting. It's another docy." And maybe for an important meeting, it's worth it. But I've tried a lot of agents and if that's the thing they want to be proactive about, I got to be honest, some days it feels like just another dock to read.
And so, I wanted to actually think about now that we have these tools, what do we do with them that's useful? How can we start to think logically through the gaps that we're closing in agentic workflows that allow us to get real work done and then start to say, okay, what are the use cases they unlock that get stuff off our desks? And so this video is about what's actually worth building once you have the power of Claude in your pocket with dispatch and computer use and all the rest of it. It's about how you clear out the technical debt. It's about how you clear out that hum or that buzz in your head when you have an idea and you can't write it down. It's about how you make sure that you make good decisions with high-quality information when you're out and about.
Look, I've been automating systems for a long time. And one of the things you learn about automating systems is that if they don't do work while you sleep, they're probably not worth it. And that's something that I notice a lot is that the people who tend to know how to use their agents figure out how to kick off work that actually builds and compounds over time. That's a lot of what I talked about when I talk about the open brain concept, right? The idea that you want to have a database that you control that's cheap, almost nothing, almost free that you can then build on by just naturally putting in what works for you and the channel that works for you. It just starts to accrue more knowledge. It allows your agent to be more proactive. Basically, I'm in a quest to make the leverage you get from professional engineering available to all of us. And I think agents are the key. And I think tools like Dispatch give us a way to do that. And if you're wondering if Dispatch will plug into OpenBrain, of course, yes it will. I'll show you how. It's not that hard.
The reason I'm spending this video covering these use cases, these launches for Claude, is that I think this is the closest we've had to a widely available, secure, remote, always-on agent that can do real work. I want it to do work for us that matters now. And I want to make sure you understand why this all stacks up and works.
So the first building block, the first Lego brick here is scheduled tasks. It's not the same as scheduled tasks in chat. This is something that OpenClaw users really loved because it let you do stuff while you were away and the Mac menu was on. Well, Claude scheduled tasks with Claude essentially closes the same gap. You get a repository, you get a schedule, and you get a prompt. And Anthropic's infrastructure runs it, whether your laptop is on or not. Not a closet server, a controlled cloud environment with configurable network access, environment variables, and setup scripts. Basically, scheduled tasks from Claude gives you a computer in the cloud that you can message and tell it to do stuff on a schedule.
Now, I will say there's some limitations, right? They're not going to let you check every minute for scheduled tasks. If you want something that's done constantly, this is not the environment for you. But if you want something checked every hour, every two hours, every three hours, and it's just a recurring task, this is perfect. And this is something that I think really speaks to developers. In fact, that's where Anthropic went in their launch video. They talked about using this exact scheduled approach internally. By the way, Anthropic dogfoods their stuff internally obsessively. That's part of sort of how they build and how they know what to build. They said that they use this sync to maintain a Go/Python library internally. In other words, a codebase in one language that automatically stays in sync with a codebase in another language because of this scheduled task function. So that's a production workflow normally would require an engineer spending a few hours a week on a task that is important but never urgent. Exactly the kind of work that tends to fall through the cracks in most engineering organizations. And yes, you can connect it to any MCP server you've already wired into Claude, right? You can connect it to Linear, you can connect it to GitHub, you can connect it to Slack, you can connect it to OpenBrain, you can connect it to Google Drive, whatever you've already set up, and the connectors just carry. You don't have to configure them twice. And so anytime you have scheduled tasks, you can just say, "Run this script at 2 a.m. every night," and it will just do it.
So what does this look like if you're not a developer? Let me give you some ideas that actually get work off your desk. Number one, I find that I constantly need to keep up with AI news. Big surprise. It's a lot of what I have to do. It is awfully nice to wake up in the morning and have a lot of that work done because of a scheduled job. And that is something that I have had to hack around on for a while. And it is a whole lot simpler when Anthropic just launches a basic primitive that says just schedule it out. And it's a full job that can run and it can stick it into your MCP server and put it in your OpenBrain and it's just done. And then I can get a parsed version, I can get a research version, I can get a condensed version, I can get whatever I want, right? I can I can edit the schedule if I want it, etc. And so if you're trying to keep up on AI, that's an example of something that like you just don't have to put as much work into reading the news because the news gets fed to you. Another great example of what you can do with this, let's say, look, aviation prices are up. You're looking for good deals. You don't want to have to scan the airline prices all the time. You get the idea, right? You can run that schedule every hour. You can come back and it will just pop in and alert you when it goes below a certain threshold for the route that you're looking at. Anything that happens in time, this can take care of for you. Here's another one. There are some places that don't let you put auto payments on a schedule. You have to remember to pay your bills on time. Why not just give yourself a reminder and it pops up for you so you never forget to pay the bill? I think that one of the biggest challenges in 2026 is recognizing how much of our lives is actually agent-shaped. There's so many problems that we can start to solve if we have these primitives in place. And so what I call a primitive is basically just a building block that you can build a lot of the rest of the work you want to do on. In this case, the building block is just native scheduling with Claude. But look how it comes together.
Number two is persistence. So if we talk about scheduling as a primitive and what Claude released there in the last few days, the second thing they released, I mentioned this at the top is dispatch. So OpenClaw also has persistence. Claude is very deliberately copying OpenClaw here. It's not a secret. OpenClaw remembered your context. It carried work across sessions. You didn't start over every time you opened the app. That continuity was a feature that felt really obvious once you experienced it and that every traditional chatbot interface lacked and people responded to this, right? This is part of what people want is an always-on agent. Dispatch closes that gap for Anthropic, but not in the way a lot of the coverage has talked about. So the surface-level description has been, look, it's just persistent chat for your phone. It's not that hard. You just chat and it's one long context window. But that undersells what is really going on.
When you pair your phone with Claude desktop via QR code, you don't get a single thread that runs a task. You're actually getting an orchestration layer. So from one conversation on your phone, you can spawn and manage multiple Claude work task sessions running simultaneously on your desktop. Each session runs independently, its own context, its own file access, its own connectors. Your phone is the command surface for all this and your desktop is the execution surface. The sessions can run in parallel. That's not remote control. That's actually a dispatch layer in the literal sense of the word. That's probably why they named it dispatch. You're dispatching work to parallel agents from a mobile interface wherever you happen to be. And what that means is you are able to get all the work you can envision for Claude co-work done without having to be at the computer anymore. And let me tell you, for some of us, especially if you're a parent, if you're out and about a lot, it is awfully nice to not be tethered to the desk.
I actually love this example from Pavle Hurin. He's a product manager who ran dispatch for 48 consecutive hours on work and he documented sort of the behavioral shift and he's a parent too, right? He went to a kid's bounce house and the work ran without him from the sidelines. He had his phone in his hand and he could direct multiple rounds of iteration on something he was working on as he watched the kids bounce around. I relate to that, right? Like I've been to that bounce house. He could do competitor analysis. He could draft sort of messaging to stakeholders and he could do all of this running in parallel across multiple co-work instances and he actually tracked how much time he spent and it turns out that over a couple of days he spent roughly 25 minutes actually entering commands in and then Claude just executed in parallel over multiple hours of work and he didn't have to fill in the dead time at the desk while Claude was tokenizing. Have you ever been the guy at the desk? You're like, "Oh, well, Claude's tokenizing. I don't know what am I going to do." And so, you don't have to do that anymore. You can do what you need to do and go about your day and just check in with Claude when you need to. Really, what's going on is that this is parallel asynchronization from your pocket.
Now, the constraint here is really obvious. Unlike the cloud example, if you want this to work, you got to have your desktop up and it's got to be on. There's no substitute for that. Now, I suspect very, very strongly all the indicators are written on the wall. Claude is going to fix that. Claude is going to make it so you don't have to have your laptop always on or it could wake itself up or whatever it needs to do. It's not here yet, but I bet you it's coming soon. And there's some other minor teething issues I expect to work through here. Every subtask that Dispatch spawns requests folder access on your desktop individually that you have to approve. There's no bulk approval right now. You can't attach files from the phone yet. You can't even receive output files back directly. The workaround is syncing your co-work instance to your Google Drive or your Dropbox so that it dumps the files in both directions automatically. And yes, you can do that and I can show you how. Complex multi-app tasks are going to succeed roughly half the time based on early testing. So, think about it as big multi-step tasks and you're getting about a 50% success rate. But if you're out and about all day, maybe you don't mind that because hey, you flip a coin, you get the success. If it doesn't work, you just try a fresh instance and maybe that one works. That's probably why Claude has labeled this research preview.
But I think the unlock you get from being freed from the desk is really huge because fundamentally it moves you into the correct pattern for agents. You are in a management pattern. When a manager is truly managing a person, do they sit there and look over their shoulder? I mean, I've had managers that do that. I didn't like those managers. Most people don't. You want the manager to go about their day and let you do your thing. That's the mode we need to get into in 2026 with agents. And that is why I keep harping on the fact that we need to expect our agents to lift work off the desk and not just create more busy work for us in the name of AI agents. And so if you're thinking about using dispatch, don't think of using dispatch just to make up work for dispatch. Instead, look at the kind of work that you would love to do in the background while you have other stuff going and like you got to cook dinner, you got to pick up the kids from school, whatever it is, and you want to spin up a Claude co-work task and just sort of keep an eye on it on the phone. That is the perfect example.
Now, anything co-work can do, and that is a broad spectrum, dispatch can do as well. And so, if you tell Dispatch, hey, I'm trying to do some deep research. I'm preparing to write a complex research report at my desk. I want to think about the way our brains are changing as we relate to artificial intelligence, which is the topic I've been thinking about. You can just tell Dispatch to dig into the research and Claude Co-work will spin that up and do all of that and have that ready for you by the time you get to the desk. If you want, for example, to have Claude Co-work dig in and start to build something for you, start to code something for you. So, if you can get to the desk and review the code block and review the prototype and see how it's going, you can kick that off as well. This is the kind of work I want to encourage you to think about. And by the way, don't be afraid of the code. If you haven't coded before, trust me, I have lots of guides on this. It is not scary. You can actually build interesting tools now straight from co-work. In fact, I know someone who never coded before in her life who has built an entire complete calendar app that meets her needs and that nobody else has been able to code for her for 20 years. And she did it in like two weeks with no coding experience. And I don't say that as an exception. She's great, but like everybody can do that now. And so this is one of those things where I want to keep emphasizing these agents are really, really good. If you have clarity of intent and you want to get work off your desk and say, "I really need this. This is going to save me hours a week. Please make it right." That is the kind of urgency I want us to bring to agents. And then dispatch frees us to be the manager and say, "So, start to work on that calendar app. Start to work on whatever the thing is that is your pet project. Let me come back to it later." Right? I'll come back and check on you in an hour. And this is more and more the pattern we're going to see in 2026.
Okay. And the last one, the thing that I think is the best is what Claude saved for last. What about the apps that don't have MCP servers? What about the apps that don't have connectors? I think this is the most consequential piece for builders and I think it's really, really big that we understand what this means. So I love that Claude reaches for MCP connectors first. It's very, very well-known. They prefer it. They love it. They encourage you to use it. There's a reason I put OpenBrain into MCP. There's a reason everybody puts things into MCP. There is a reason that when Google launches their products, they talk about MCPs that work with Claude. MCP is the universal USB of the AI age. But despite that, not everybody puts their stuff into MCP. It's just a reality. The web is vast. There are many, many tools. There's just never going to be a world where we have close to 100% coverage. There's never going to be a world where we have 50% coverage with MCP servers, right? More than half the world is not accessible to agents easily. And we're talking just about the limited world of the screen. Still, more than half of that is not accessible to agents.
And so what Claude did here is it took the blinders off and it basically said, "You know what? We're going to make sure that remotely through co-work tool on your computer and that includes the web browser and you can do any kind of remote work that involves keyboard and mouse clicking." That's right. You can be, you know, 10 miles away watching the kids at the bounce house and your computer is going to be typing away and your mouse is going to be moving and it's not going to be haunted and it's not going to be a horror movie. It's just going to be Claude doing his job because you asked it to via dispatch. And this is really a case where you start to solve business problems, where you start to get work off the desk. Think about the old instance of Jira. Maybe it's not up to date. Maybe it doesn't have a good MCP. Think about a bespoke ERP screen. Think about an old SAP instance. Think about something that is not ever, ever, ever going to be easy to get data out of. And maybe you don't have time to build the full connector with the engineering team because you can't get on their backlog, but you have a process that has to run and you have to get it into a spreadsheet and then you have to send it to the accountant. I've been there, right? I've been the person who had to go through the agonizing process of checking two or three different antique sites for data that I then manually put into a spreadsheet and it took me like half the day. There is so much office work that falls into that category. And essentially what Claude said is you're the manager now. Just send the task via dispatch. Walk away and let Claude manage it for you. Let Claude get that out for you. And that's why I keep harping on like think like a manager. Get work off your desk. Don't be satisfied with what I would call pseudo work. So much of the demos we see brag on pseudo work. I was looking at another one today. It's like so obsessed with proactive briefing. So obsessed with essentially making more text for me to read. Stop making more text for me to read. Agents should take work off my desk. If they're not taking work off my desk, I don't want it.
And this is what makes this exciting is because this is actually a string of Lego bricks, a string of primitives you can put together that you can actually get real work done with. You can assign the task from your phone. You can use desktop apps to complete it and it can be totally done. If you're the engineer, you can schedule it in the cloud and it can just be done. And people are going to tell you at this point that the key difference between OpenClaw and Anthropic is safety, that Anthropic is the safe one, and OpenClaw is the dangerous one. I don't think that's correct. I think that the difference is the degree to which you want to self-host versus the degree to which you want a managed agent.
So, OpenClaw's infrastructure that you have to roll and run yourself. You set the server up. You configure the network. You manage the credentials. You vet the skills. You troubleshoot the websocket connections. You decide what the agent can access and what it cannot. For developers who want that control, it is a super powerful thing. For everybody else, it kind of feels like a second job. The Mac Mini in the closet does need to be maintained. And for developers who know how to do it, it's not too bad. They can get it done. I've written the guide on how to do that. You can totally do it even if you're non-technical. But I don't want to pretend it's not work.
Anthropic brings managed infrastructure. Cloud scheduled tasks run on their servers, not yours. Dispatch runs in a sandboxed environment where Claude accesses the files and the apps that you've explicitly stated that it can access with permission. Computer use is going to ask you before it touches new applications. You don't configure the network. You don't vet the skills marketplace. You don't configure the server. So much of that is abstracted away for you. And I want you to think back to the history of compute here. This is the exact same shift that happened with email where we went from self-hosted Sendmail to Gmail for those of you who have gray beards with compute where we went from rack servers to AWS with CI/CD where we went from Jenkins on your box to GitHub Actions. Every time the self-hosted version comes first and proves that the category exists. The managed version comes second and that's the thing that gets mass adoption for you. It's not because the self-hosted version was wrong, quote unquote. It's just that that's the way you learn if something works and if people want it. And once you prove that, it's relatively easy to roll out the managed version to a wider audience, which is exactly what Peter Steinberger is working to do at OpenAI. I guarantee you, they are looking at what Claude is launching over at OpenAI and they're like, "We can't wait. We're going to launch soon."
And look, I'll be honest, the ceiling is different here with agents, right? OpenClaw is going to give you more raw freedom. It is the dangerously skip permissions version of agents and Anthropic is going to rein it in a little bit. Anthropic's stack runs on cloud. You can't just swap any LLM in. It reaches what your connectors and computer use can touch. If you want to wire in your local Llama on-prem, I'm not going to pretend this is for you. But for most of us, for the vast majority of people who are passionately interested in agents, and I get asked about this all the time, when is there an agent? When is there an OpenClaw? Is OpenClaw safe for me? This is the safe OpenClaw, right? This is safe enough. It's useful enough. It does enough that it works. And that's why I keep emphasizing tasks that get work off our desk. I want us to think about that. I want you to comment on this video and tell me the tasks that get work off the desk versus the fake work, the stuff that works in demos that doesn't actually get work off the desk.
I'll give you another one from trips. I talked earlier about the fact that scanning for airline prices and saving real money, which can save you the cost of the closet subscription, hey, that gets work off the desk. You know what doesn't get work off the desk? Every single time they demo the special travel, trip preparation, whatever. Look, I do spend a little bit of time planning a trip, but I do not understand why that has to be the showcase demo for every single app that is agentic. We don't all plan trips quite that much.
So before I close this video, I want to give you the framework to start to figure out how to get stuff off your desk and give it to agents in the way that we need to in 2026 with a management mindset. The first principle I want to teach you here is that you need to find something that is buzzing in the back of your brain that you need to close. Look for it as a commitment loop that you need to close. Every promise you make and people, the more responsible you are, the more promises you have to write, whether you're promising in email, whether you're promising in Slack, whether you're promising in WhatsApp, it's an open loop. You told the client you'd send the revised scope by Thursday. You signed the memorandum of understanding by Wednesday. You told your team that you'd get the minutes from the last meeting to them by tomorrow. Whatever it is, you made a promise, you need to keep it, and it's sitting in your head. That is the kind of thing that you need to be delegating to agents. And especially if you are someone who has lots and lots of these open promises at once, you got to close them. You got to give more of them to agents and let those agents do the work and carry the mental load forward and close it. And if you're going to tell me, Nate, I can't do it because the agent's quality isn't good enough, I'm going to say, that's a skill issue. That's a prompting issue. And that is something we can fix. That is something we can fix because you can learn to give the agent what it needs either through the system instructions, through the context layer that you're giving it, through the prompt that you're giving it to get a quality result. It is not impossible in 2026. I do not want to hear, "I can't get a good report out of AI." I don't want to hear, "I can't get a good draft of the memorandum of understanding." "I can't get good meeting minutes out." No, no, no, no, no. You can. It's a skill issue and you can learn and I have a guide for that. You can figure it out.
I'll give you another category that I think is really important to get work off the desk. If you are trying to make a decision, so often we come against these decisions in business and we don't have time to consider the information involved and make a good choice. We're running into the meeting, we're late, we didn't read the docs in advance. How many of us read the docs in advance? Let's just be honest. And we're like, "Okay, what do I have? What do I figure out?" And do you know how many people sit there and they're stuffing the docs into ChatGPT and they're like, "Tell me what to think. Tell me what to do. Tell me what to decide." Please don't do that. You have time to figure this out. You can run a scheduled task for this. You can tell your dispatch in the morning on the way to work to help you develop an understanding of the subject area. You can figure out your opinion by bouncing it off of an AI, even if you're on the phone and getting data back. And by the way, there's a way to do this badly and well. The bad way to do it is to try and use the AI to confirm your own opinion. The good way to do it is to actually push and say, "I need more data and more information than I would normally use to make this decision." Typically, if I'm running into a meeting and I'm late, I'm making the decision with 30% of the available information or less. Why not make it with 70%? Why not go fishing for data that would help you make a better choice? Maybe it's using computer use and finding the dashboard that you normally wouldn't have time to check. Maybe it's digging up the Excel spreadsheet you normally wouldn't have time to grabbing and pulling it in. The more you start to use the agentic tools at your disposal to pull in that data, the better off you're going to be.
Compound signal detection. I have I have emphasized this with OpenBrain. I'm just going to come back to it again. If you have attached your OpenBrain, you have so much value here because you are now on your phone with your dispatch and your Claude knows OpenBrain and can talk to OpenBrain via the MCP server and you can start to benefit from compounding signal detection across weeks wherever you happen to be. And so your agent might come back and say, "Hey, you ran a scheduled task. I noticed that your competitor is hiring aggressively around payments. This is something we talked about when we talked about corporate strategy three weeks ago. I think you should do X or Y." Or your agent might be doing research and say, "Hey, a competitor filed two patents. The patents are around cross-border settlement. I'm connecting this with their hiring patterns. Now we're seeing their cross-border strategy emerge," etc., etc. Essentially, if you have dispatch and you have OpenBrain, you have a proactive way to figure out patterns over time wherever you happen to be. On a schedule, not on a schedule, using computer use, not using computer use. It takes the agent from being reactive to what you stick into dispatch to being something where you can proactively pop up, see an email from your agent, and say, "Ah, that's something I want to pay attention to."
And this one is for engineers. We loved doing overnight engineering work, jobs that would run all night when I was at Amazon, because it's work that you can do and it cleans stuff up while you sleep. It's perfect. Hopefully nobody gets woken up. I don't mean run the linter, run an analysis of your code file. I mean migrate this dependency. I mean please improve test coverage, write test coverage to cover 80% of use cases. I mean refactor the authentication layer we can never get to stuff like that. And if you think this is too technical for you, I'm going to remind you there are people who are coding whole apps who have never coded before. If you have clarity of intent of what you want to get done, the technical piece is not an obstacle. In fact, you can delegate a lot of it in 2026 and you can get a surprisingly useful result. I'm not going to pretend that you can vibe code Amazon.com or you can vibe code Salesforce and it will be exactly the same. But I will say there are people who have vibecoded their own CRM and are perfectly happy at startups. So the distance is maybe not as much as you think.
The pattern across all of these cases that I have called out to you in the last few minutes is consistent. It is work off the desk. It is the agent takes something off of your plate. That is the meta-skill that I want you to learn. Look at these primitives as free leverage. This is leverage you can use in 2026 to get more of what you want done off your plate. That is what we should expect from our agents.
I want to close with a reflection that I think is a big piece of this year that I don't see talked about anywhere. People talk about agents like the next best thing since sliced bread. They put all these demos out that are basically fake work, which we've talked about extensively in this video. But what we don't talk about is the fact that it is hard for us as humans to trust that the computer is doing work when we aren't looking at it. I have this problem, too. I walk away. I use my dispatch and I'm itchy. I'm like, I want to go back. I want to make sure it's working. Is it really working? We're going to need to learn to shift that. We're going to need to learn to untether. We're going to need to learn to trust that the agent is doing the work when we walk away. That is one of the big shifts I see in the next half of the year. And it's going to be the people who are able to walk away, who are able to touch grass, who are able to see the kids at the bouncy castle, who are able to reset, who can make good decisions about what they want to give their agents, and who ultimately get the leverage back that this whole AI thing is all about. Because AI is really not about the fake work. And people tell me it's a bubble. It's a bubble. I'm like, no, it's not a bubble. The demand for tokens is outstripping supply and we're building it as fast as we can. This is not a bubble problem. It is a how do we use it intelligently problem and that is on us and we have got to figure out how to change our skill set so we are more and more and more managers as we move into the second half of this year and that is the story on dispatch that no one is telling you. Cheers.