The Sequence Opinion #864: Every AI Agent Needs a Computer
The raise of agentic sandboxes.
The next phase of AI agents will not be defined only by better models, longer context windows, or more elegant tool-calling APIs, but by something much more primitive: access to a computer. An agent that can only emit tokens is a brilliant brain in a jar; an agent with a filesystem, terminal, browser, network, package manager, credentials, memory, and guardrails becomes a worker inside a real execution environment. This is the core thesis: every serious AI agent needs a computer, not metaphorically, but architecturally. It needs a safe, isolated, programmable space where it can write code, run commands, inspect outputs, manipulate files, browse the web, recover from errors, and iterate through the same feedback loops that make software useful. The emerging market for micro-containers, sandboxes, browser runtimes, and agent workspaces is really the market for giving intelligence a body.