An AI Agent Is AI With Its Own Computer
The chatbot handles the flexible reasoning, while a separate, reliable computer handles the exact work.
Stratechery · Apps, Agents, and Aggregation | Stratechery by Ben ThompsonAn AI agent is not a smarter chatbot. It is a chatbot that has been given its own computer to use.
That is the distinction drawn by technology writer Ben Thompson, who has tracked artificial intelligence since ChatGPT first gained outside tools in 2023. The intelligence does the reasoning, and the computer does the computing.
Language models like those behind ChatGPT and Claude work by predicting likely responses. That makes them good at flexible conversation and reasoning, but unreliable at the precise, repeatable work computers do well.
Traditional software is the opposite. It is deterministic, meaning it follows exact instructions and gets the same result every time. An agent combines the two, letting the language model direct a regular computer to perform computer tasks, while also writing things down so it can keep track of multi-step work.
Most people cannot, or will not, set up and maintain such a machine themselves. Thompson argues that to make agents useful to ordinary users, a company must provide not only the agent but also the computer it runs on.
That is why he focused on Meta's move with Muse, Meta's AI assistant. Thompson notes that Meta was provisioning each user in the United States with a virtual machine, which is a full computer running in Meta's data center rather than a box on a desk, with two processor cores, 8 GB of memory, and 8 GB of storage.
With that setup, you do not need a more brilliant AI. You need adequate AI paired with adequate computing, so the AI can operate apps and websites on your behalf the way you would yourself.
Questions people ask
What is a virtual machine?
A virtual machine is software that acts like a complete physical computer, with its own processor, memory, and storage. It runs inside a company's data center, so you can use its power over the internet without owning or maintaining the hardware.
Why can't the AI just do the computing itself?
Language models predict likely words, which makes them good at conversation and reasoning but bad at exact calculation and repeatability. A conventional computer follows precise instructions, so pairing the two lets each do what it does best.
Why does an agent need storage and memory?
Because the agent can keep files and a record of what it did on that computer. That stored record lets it carry out multi-step jobs and pick up where it left off.