I have a test for every AI agent demo I see: does it take work off my desk, or does it put more work on it?
Most fail. They want to give me a “proactive briefing.” Congrats, you’ve created another doc for me to read. Another summary I didn’t ask for. That’s not help. That’s homework with better typography.
Anthropic just shipped three features that turn Claude into an agent that actually passes the test: Scheduled Tasks, Dispatch, and Computer Use. I’ve been running all three for some time now, and the difference between these and the typical “AI agent” demo is that when I walk away, work actually gets done. Not a draft for me to review. The thing itself.
What Claude agents are actually doing for me
Here’s what my mornings look like now. Before I sit down at my desk, a Claude scheduled task has already pulled my email, outlined what I need to respond to, and time-blocked my calendar around my meetings using the Fantastical MCP. I sit down and it’s all there. No triage. No “let me figure out what today looks like.” It’s figured out. (I wrote about my full Claude Code setup a while back.)
I have AI news aggregation running on a schedule so I’m not spending the first hour of my day in an RSS reader. Weekly pulls from GA4, GSC, and Ahrefs feed directly into my reporting. My Uncommon Stack newsletter has a scheduled Cowork prompt that asks me questions every week so I keep the same format and don’t have to stare at a blank page.
Some bills don’t let you set up auto-pay (why is this still a thing in 2026?). So I have Claude remind me. And it doesn’t just remind me once. It asks me later in the day if I paid it. It keeps asking until I confirm I did or didn’t. That’s not a notification. That’s an agent that closes the loop.
I’m planning the annual baseball trip with my son. Every year we pick a city, catch a game, eat too much, the whole thing. I have Claude searching flights on a schedule and it’ll ping me when the price hits the number I want. Anything that happens in time, this can handle for you.
None of these examples generated a report for me to read. None of them produced a summary I had to act on. The work just got done. That’s the difference between an agent and a chatbot wearing a trench coat.
Our lives are agent-shaped. We just didn’t have the primitives to prove it until now.
Dispatch is Open Claw for the rest of us
Here’s the part most coverage undersells. From one conversation on your phone, you can run multiple Cowork sessions simultaneously on your desktop. Each session runs independently with its own context, its own file access, its own connectors. Your phone is the command surface. Your desktop is the execution surface. The sessions run in parallel.
That distinction matters more than it sounds. If you’ve ever sat at your desk while Claude is tokenizing, you know the feeling. You’re just… waiting. Staring at a progress indicator like it’s a microwave heating up yesterday’s coffee. Maybe you check Twitter. Maybe you open a tab you shouldn’t. (We’ve all ended up on Reddit at 2pm because Claude was thinking.)
Dispatch means you don’t have to do that anymore. You dispatch the work, you go do something else, you come back when it’s done.
For big multi-step tasks, I’m getting about a 50% success rate right now. That sounds bad until you realize you’re running these from the sidelines of your son’s baseball game. Flip a coin, get the work done for free. If it doesn’t land, spin up a fresh instance. That’s why it’s still labeled “research preview,” but the pattern is right.
It’s like being a manager. You tell Claude what to do via Dispatch and you go about your day. Anything Cowork can do, Dispatch can do too. You just don’t have to be at the desk for it.
Computer Use closes the last gap
MCP is the universal USB of the AI age. (I called it the jQuery of AI agents last year. The label stuck.) I use it for everything I can. But not every app has an MCP server, and they never will. Half the tools we use at work were built before MCP was a twinkle in Anthropic’s eye. Claude Computer Use means it can open apps, click through screens, and work with tools that have no API. Old Jira instances. Bespoke ERP screens. That one internal dashboard held together by prayers and a single engineer who left in 2019.
This is where the whole stack comes together. Scheduling, persistence, and computer use stacking into something that didn’t exist before: a real always-on agent that can touch anything on your machine. Yes, your mouse will move by itself. No, your house is not haunted. Probably.
The “is this safe?” question
People will tell you the difference between Open Claw and Claude is safety. That’s not quite right. The difference is how much infrastructure you want to babysit.
Open Claw is self-hosted. You set up the server. You configure the network. You manage the credentials. You vet the skills. You troubleshoot the websocket connections at 11pm on a Tuesday when you should be watching baseball. For developers who want that control, it’s great. For everybody else, it’s a second job.
Claude’s stack is managed. Scheduled tasks run on Anthropic’s servers, not your closet Mac Mini. Dispatch runs in a sandbox where Claude only touches what you’ve explicitly approved. Computer Use asks before it opens new apps. You don’t configure the network. You don’t vet the skills marketplace. You don’t maintain the server.
This is the same pattern as every infrastructure shift ever. Self-hosted email to Gmail. Rack servers to AWS. Jenkins on your box to GitHub Actions. The self-hosted version proves the category exists. The managed version is the one your mom can use. (No offense, Mom.)
Open Claw gives you more raw freedom. Claude gives you guardrails that don’t require a systems engineering degree to set up. For most people who keep asking me “is Open Claw safe?” the answer is: Claude’s version is safe enough, useful enough, and you won’t have to debug websockets to get it running.
How to find the work that’s agent-shaped
Here’s the framework I keep coming back to: look for open commitment loops.
Every promise you make is a loop. You told the client you’d send the revised scope by Thursday. You told your team you’d get the meeting minutes to them by tomorrow. You said you’d check on that invoice. Every one of those is sitting in the back of your brain taking up space, humming like a refrigerator you can’t turn off.
The more responsible you are, the more loops you’re carrying. (Congrats on being reliable. Your reward is anxiety.)
These are the things you should be handing to agents. Not because the agent writes a perfect email on the first try, but because the agent can close the loop. Draft the scope. Pull the minutes. Remind you about the invoice and then nag you until you confirm it’s paid. That bill reminder I mentioned? That’s a commitment loop getting closed without me having to remember it exists.
The other category: decisions you’re about to make with 30% of the information. You’re running into a meeting late. You didn’t read the docs. (Nobody read the docs. Let’s stop pretending.) Instead of stuffing everything into a chat window and asking “tell me what to think,” use a scheduled task or Dispatch to pull the data beforehand. Not to tell you what to decide. To make sure you’re deciding with 70% of the information instead of 30%. That’s a different outcome.
Agents need to take work off your desk
I keep seeing demos where an agent’s big trick is generating a “proactive briefing.” A doc. A summary. Another thing I have to read before I can do the thing I was already going to do. That’s not taking work off my desk. That’s just AI-generated busywork. We had enough human-generated busywork already, thanks.
Stop giving me more to read. Start finishing things.
The agents that matter in 2026 close loops. They pay the bill. They find the flight. They block the calendar. They pull the data and put it in the report. Not a summary of the data. The report itself. The work is done when I come back, not waiting for my approval to start.
That’s the bar. If your agent isn’t clearing work off your desk, it’s just a chatbot with a cron job.