Delegate All Your Grunt Work to AI Agents
A year ago AI couldn't reliably operate a computer. Now it can, and it already wrote code. Three new agent products prove the change. Have you delegated all of your grunt work yet?
Something material changed in the past year, and I want to make sure you noticed.
A year ago, AI was already good at writing code, but it wasn’t good at operating a computer. Now it is. That means it is now possible, and rational, to delegate all of your computer grunt work to AI agents.
So here is my question: have you delegated everything yet?
The new agents
In the past three months, three of the biggest AI labs each shipped a new kind of product: an agent with its own computer, which keeps working while you do something else.
- Grok Bot, from xAI, now called SpaceXAI, opened in early beta on August 11. Each bot works on a persistent cloud computer with a browser, filesystem and terminal, keeps your logins, and keeps running after you close your laptop (xAI docs, xAI guide, VentureBeat).
- Dots, from OpenAI, launched at DevDay on September 29. They are always-on agents in ChatGPT that launched on GPT-6 Astra, and each one works from its own cloud computer (OpenAI, SiliconANGLE).
- Claude Duet, from Anthropic, opened as a research preview on October 21. It is powered by Claude Opus 5.5, “works on its own computer in the cloud,” and asks for your approval before sensitive actions like sending messages or paying (Anthropic, CNBC).
These are big improvements on what came before. They run on their own machines, work in your apps, and come back to you when they need a decision.
A year ago it couldn’t, now it can
The products are only possible because the models got good at operating a computer.
OSWorld tests an agent on ordinary tasks on a real computer, using the screen, mouse and keyboard. When its authors published it, people completed 72.36% of its tasks (OSWorld). Its team now runs OSWorld-Verified, a corrected version of the benchmark, and lists verified results apart from self-reported ones (OSWorld-Verified).
Compare like with like on that leaderboard: a general model, one attempt per task, up to 100 steps.
- A year ago: the best was Claude Sonnet 4.5, at 62.88%, in a result dated October 31, 2025.
- Now: the best is Claude Fable 5, at 85.96%, in a result dated August 1, 2026.
A year ago the best general model was below the human rate. Now it is well above it.
And AI was already good at coding, and it got much better. METR, an independent research group, measures how long a software task can be, counted in the time it takes a human expert, before an agent fails half the time. GPT-5, released in August 2025, could handle tasks of about 3.4 hours. Claude Opus 4.6, released in February 2026, could handle about 12 hours, and the newest model METR measured went past the 16 hours its tasks can reliably measure (METR).
One app left
If you delegate all of your grunt work, you are left with one primary app: a conversation with an AI agent.
That single interface can do everything. You no longer need to learn a million different user interfaces, web apps and command-line tools. You say what you want, and the agent operates them for you. The software is still there. You just don’t have to be the one who operates it. That is also why a new market is opening for software built for agents, and why chat systems for AI agents matter so much.
Your mind back
Grunt work doesn’t only take time. It takes attention. Every form, every dashboard, every settings page and every export takes a piece of your mind.
Hand that to the computer, and a large part of your mind is free for what matters most: deciding what to do, judging what’s good, and working on the problems only you can work on. That is a huge unlock. It is real leverage.
The window is closing
Right now, the people who delegate everything are ahead of the curve.
That won’t last. In a year, doing this will no longer put you ahead. It will be what everyone does. Eventually, the people who don’t do it will fall behind.
So the leverage is outsized only now, before your competitors catch on. I already made the general case in Delegate Everything You Can to AI Agents. This is the reminder that the tools have arrived, and the clock is running.
Have you delegated all of your grunt work yet?