// Sandbox for agents
A workspace for agents.
A sandbox is an isolated workspace an agent can use, created and managed through an API. The agent can run code, work with files, install libraries, test applications, and use a browser when needed. FAR Labs provides the execution environment; your application controls the agent's logic.
What it looks like in practice
The user asks: “Analyze sales and create an Excel file with charts.” Here is what happens next.
Plan and code
The model creates a plan and writes Python code.
Create the sandbox
Your application creates a sandbox through the API and places the source files inside it.
The code runs
Inside the sandbox it reads tables, calculates metrics, and generates charts.
Feedback loop
Your application sends the result or error back to the model; the model fixes the code if needed.
The user gets the file
The finished file is returned to the user.
Delete or retain
The environment is deleted — or retained so work can resume later.
Two product options, plus storage
Inference provides decision-making and code generation; the sandbox provides execution and working state. Both can be offered together — model inference and state storage are billed separately. Sandboxes primarily use CPU and RAM; GPUs are needed only for models and specialized computation.
Full environment
A complete workspace for agents that need code, files, packages, tools, and browser access.
Code execution session
A simplified execution session for agents that only need code, files, and results.
Storage
Persist working state when agents need to pause, resume, or return files later.
Give your agents somewhere to work.
Pair sandboxes with inference through the same platform.