OpenShell DeepAgent is a reference implementation of a general-purpose coding agent that writes and runs code inside an NVIDIA OpenShell sandbox. The agent itself is built with LangChain's Deep Agents harness — providing task planning, a virtual file system, subagent spawning, and long-term memory on top of the LangGraph runtime — and is driven by NVIDIA's Nemotron Super 3 model served through NVIDIA NIM. Its purpose is to demonstrate running an autonomous coding agent in an isolated, policy-governed Linux environment where security is enforced outside the agent rather than through prompt instructions alone.
OpenShell acts like a browser's security model for agent code execution: an on-prem sandbox that lets the agent write and run code while enforcing policies over filesystem, network, and process permissions. The example shows this in action — a network policy blocks the agent from making outbound requests regardless of what it is told to do. Policies are defined in YAML and can be hot-reloaded without restarting the sandbox.
Architecturally, the Deep Agent exposes tools (execute, writefile, readfile, edit_file, glob, grep, ls) that route code execution to an OpenShellBackend over gRPC while a separate filesystem backend keeps memory and skills on local disk that persists across restarts. Code runs in a sandbox container managed by an OpenShell gateway (k3s inside Docker). The backend is designed as a drop-in replacement for a Modal-based backend — swapping ModalBackend for OpenShellBackend leaves memory, skills, and subagents unchanged. It runs locally with no cloud dependency beyond the hosted model, and is developed and driven through the LangGraph dev server and Studio UI.
