Coding agent

Use this pattern when an agent needs to inspect or modify a project but should not run directly in the host process.

Prepare a workspace

Create a narrow project directory for the task. Avoid mounting $HOME, SSH agent sockets, cloud credentials, browser profiles, or unrelated repositories.

mkdir -p /tmp/mvm-agent-work
cp -R ./src /tmp/mvm-agent-work/src

Build the runtime

Prefer a Nix flake that declares the tools the agent needs:

mvmctl build --flake .

For local development through this repository, the builder VM is the Linux build boundary. The runtime guest later boots the built artifact.

Run the task

mvmctl up --flake . --name coding-agent
mvmctl exec coding-agent -- bash -lc 'cd /work && python task.py'

Use file transfer or a narrow mount for input/output. Keep generated patches and logs outside broad host write access.

Network policy

Start with no egress for code analysis. If the task needs registries or model APIs, add only those destinations and record why they are needed.

Persist or discard state

Use cold mode only when the agent needs to resume an environment with installed packages, caches, or intermediate files.

mvmctl pause coding-agent
mvmctl resume coding-agent

Use mvmctl down and cleanup commands when the task is complete.

Security checklist

• Do not pass long-lived credentials as plaintext environment variables.
• Keep host mounts narrow and read-only unless write access is required.
• Treat command output and logs as sensitive.
• Keep network allowlists task-specific.
• Save snapshots only when the task requires recoverable state.