SWE-ReX provides a runtime layer for agents that need to execute commands inside sandboxed shell environments without baking infrastructure details into the agent loop. The operator wires an agent to SWE-ReX, chooses a local or remote backend, and lets the agent run commands, read outputs, handle exit codes, and keep multiple interactive sessions alive.

## When to use it

Use this instead of direct shell access when an agent needs repeatable execution across local machines, Docker containers, AWS or Modal-style remote machines, or many benchmark/task sandboxes in parallel. It is a good fit for coding-agent evaluation, SWE-bench-style runs, multi-session debugging, and infrastructure where the same agent code should work across different execution backends.

## Scope boundary

This is a sandboxed execution runtime for AI-agent command loops, not a full coding agent, benchmark, cloud IDE, or generic Python library listing. The skill boundary is: create a runtime, run and observe shell sessions, manage parallel or interactive command execution, and return structured results to the caller. The agent still owns planning, code edits, evaluation criteria, and safety policy.

## Operator workflow

1. Install the exact Python package and optional backend extras needed by the target environment.
2. Select the backend for local, containerized, or remote execution and keep credentials outside prompts.
3. Give the agent a narrow task, workspace, timeout, and allowed command policy.
4. Run shell sessions through SWE-ReX so command completion, stdout, stderr, and exit codes are returned consistently.
5. Use separate sessions for interactive tools such as debuggers or Python shells, and review artifacts before merging or publishing agent output.