Filesystem
The agent’s most common environment: files it can read and write.
The directory tree an agent is given access to. Reading, searching, editing and creating files is the basis of most coding agents.
Filesystem access is a large capability. Restrict the agent to a working directory, and keep it away from credentials, home directories and anything outside the project.
Compared with a database tool, a filesystem is unbeatable for agent work: it is inspectable, diffable, greppable, version-controlled, and it survives the session. That is why the most capable coding agents converge on "here is a directory, do what you need" rather than a curated set of queries.
Files as working memory
An agent needs to track findings across a long task.
Write them to a file and read it back when needed. The file is durable across compaction, inspectable by a human, and costs tokens only when read — which makes it strictly better than holding the same notes in a context window that is filling up.
- docsClaude Code overview ↗
A first-party description of the surfaces an agentic coding tool ships on, including parallel sessions and long-running tasks. Not evidence that it improves productivity.
- blogBuilding agents with the Claude Agent SDK ↗
States the agent loop plainly — gather context, take action, verify work, repeat — and shows subagents and compaction as context-isolation tools.
Environment
The world outside the model that tools act upon.
Sandbox
An isolated environment where an agent can act without risk to real systems.
Permission Request
The agent pausing to ask before doing something risky.
Agent
A model given tools, a goal and permission to keep going on its own.