The directory tree an agent reads, writes, and executes within — the default environment for coding agents.
The filesystem is the tree of files and directories an AI agent reads from, writes to, and executes within — the default environment for a coding agent. It contains everything the agent works on: source code, build scripts, configuration files, AGENTS.md instruction files, and tool definitions. When a harness starts an agent "in your project," it is pointing the agent at a filesystem as its environment.
Agents interact with the filesystem exclusively through tool calls and tool results. A Read tool exposes file contents; Write and Edit tools modify files; Execute tools run scripts or commands. The filesystem is stateful in the traditional sense — changes persist across tool calls and remain visible to subsequent operations within the session.
The filesystem as an environment introduces scope risks. An agent with write access to a broad filesystem can modify or delete files beyond its intended scope. Conversely, an agent pointed at the wrong filesystem — a sandboxed subdirectory instead of the project root — will be unable to find or operate on the files it needs. Knowing which filesystem an agent is running against is a prerequisite for predicting what it can and cannot do.
Open questions include how to give agents reliable awareness of filesystem scope boundaries, and how to maintain filesystem-level hygiene as agent sessions accumulate temporary files, generated artifacts, and modified source files without creating state that poisons future sessions.
Signals turns a topic into a sourced research record you can inspect and rerun. Your first scan is free, and this one starts with Filesystem (AI Agent) already loaded, so edit it or scan as is.