Environment
The world an agent acts in — the files, apps, websites, or systems it can touch and change.
Think of it like
The kitchen a chef works in: the ingredients, appliances, and mess that respond to what they do.
Example
For a coding agent the environment is the repository, the shell, and the test runner; its actions edit files and its observations are the test output.
How it actually works
The environment is everything outside the model that the agent perceives and affects. Its properties shape everything: does it give clear feedback, is it safe to fail in, can actions be undone? A sandboxed environment lets an agent try, break things, and learn cheaply; a live production environment turns every mistake real. Designing the environment is often as important as the agent itself.
For product teams
Where an agent runs decides how safely it can be allowed to fail.
For engineers
The external system an agent observes and acts on, defining feedback signals, state, and blast radius.
Related
- Sandbox Execution — A safely-walled environment for risky actions.
- Action Space — What the agent can do within it.
- Observation — What it perceives back from it.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome