Agent Workspace
The agent workspace gives an LLM a real Unix environment it can navigate, run code in, and install packages into — without any access to the host machine.
Instead of a list of tool call schemas, the agent gets a shell. It uses cat, echo, ls, and shell scripts — primitives it already knows. Tool calls map to filesystem operations: read a file to observe state, write a file to trigger an action.
What’s being built
Section titled “What’s being built”agentFS— a virtual fileserver exposing task state, output streams, and control files- Typed NDJSON output streams for structured agent responses
- Modal TTY separating agent I/O from user I/O
- Session checkpointing and multi-agent coordination via shared namespaces
In the meantime
Section titled “In the meantime”A fishbowl instance already gives you a sandboxed shell. instance.run('your command') executes a command and returns the output. The Hello World →, Containers →, and Custom Package → guides cover everything that is working today.
Further reading
Section titled “Further reading”- Capabilities: Agent Harness → — What the harness will do
- Containers guide → — Isolation between agent sessions
- Custom Package guide → — Expose domain-specific tools as bins