Environment
The world outside the model that tools act upon.
Everything the agent can affect through its tools — a filesystem, a database, a browser, an API. The environment is what makes the model’s actions have consequences.
Scaled autonomy scales blast radius. The same agent that saves you an afternoon on a sandbox can delete production data if the environment gives it the credentials to do so.
Environment design determines what an agent can achieve far more than model choice does. An agent with a real shell, a real test suite and a real repository outperforms a smarter model boxed into a chat window — because the environment is where verification comes from.
The missing test runner
An agent writes plausible code it cannot check.
Give it the ability to run the tests and the whole loop changes: it iterates against evidence instead of presenting a first draft. Adding one tool to the environment moved quality more than any prompt change had.
- docsComputer use ↗
Two GUI-agent patterns — structured mouse and keyboard actions, or code driving a browser — and where the runtime owns screenshots, sessions, limits and permissions.
- blogScaling managed agents: decoupling the brain from the hands ↗
Separates model, harness, durable session log and execution sandbox — the architecture that makes long-running agents survive crashes and tool boundaries.
Sandbox
An isolated environment where an agent can act without risk to real systems.
Filesystem
The agent’s most common environment: files it can read and write.
Permission Mode · Autonomy level
How much the agent may do without asking.
Tool
A function the model is allowed to call to affect the world.