Problem Statement
NemoClaw does not currently expose host-side commands to add, list, remove, or apply OpenClaw agent configuration inside a NemoClaw-managed sandbox.
DataRobot feedback: host-side agent lifecycle commands would avoid manual in-sandbox OpenClaw config mutation when setting up multi-agent workflows.
This makes dynamic multi-agent setup difficult because NemoClaw hardens OpenClaw config at runtime, and users should not need to enter the sandbox or manually mutate read-only config.
Proposed Design
Add a host-side agents command group to NemoClaw.
Possible commands:
nemoclaw <sandbox> agents list
nemoclaw <sandbox> agents add <id> --workspace <path> --model <model> --role <role>
nemoclaw <sandbox> agents remove <id>
nemoclaw <sandbox> agents apply -f agents.yaml
System behavior:
agents list should inspect the NemoClaw-managed desired state and/or OpenClaw agent config for the named sandbox.
agents add should add a new agent definition, create its workspace, assign model/tool metadata, and apply the change through the same supported lifecycle path used by manifest apply.
agents remove should remove the agent definition but should not delete workspace data unless the user explicitly requests it.
If live mutation is unsafe, these commands can stage desired-state changes and require an explicit rebuild/apply step.
Components involved:
NemoClaw owns the host-side CLI, desired-state update, workspace lifecycle, and safe apply/rebuild behavior. OpenClaw owns runtime agent behavior after the agents exist. OpenShell may be needed only if safe in-place sandbox mutation or reload primitives are required.
Alternatives Considered
Running OpenClaw agent-management commands inside the sandbox does not fit NemoClaw’s hardened config model. Rebuilding a custom sandbox image for every agent change is reliable but too heavy for normal iteration.
Category
enhancement: feature
Checklist
Problem Statement
NemoClaw does not currently expose host-side commands to add, list, remove, or apply OpenClaw agent configuration inside a NemoClaw-managed sandbox.
DataRobot feedback: host-side agent lifecycle commands would avoid manual in-sandbox OpenClaw config mutation when setting up multi-agent workflows.
This makes dynamic multi-agent setup difficult because NemoClaw hardens OpenClaw config at runtime, and users should not need to enter the sandbox or manually mutate read-only config.
Proposed Design
Add a host-side
agentscommand group to NemoClaw.Possible commands:
System behavior:
agents listshould inspect the NemoClaw-managed desired state and/or OpenClaw agent config for the named sandbox.agents addshould add a new agent definition, create its workspace, assign model/tool metadata, and apply the change through the same supported lifecycle path used by manifest apply.agents removeshould remove the agent definition but should not delete workspace data unless the user explicitly requests it.If live mutation is unsafe, these commands can stage desired-state changes and require an explicit rebuild/apply step.
Components involved:
NemoClaw owns the host-side CLI, desired-state update, workspace lifecycle, and safe apply/rebuild behavior. OpenClaw owns runtime agent behavior after the agents exist. OpenShell may be needed only if safe in-place sandbox mutation or reload primitives are required.
Alternatives Considered
Running OpenClaw agent-management commands inside the sandbox does not fit NemoClaw’s hardened config model. Rebuilding a custom sandbox image for every agent change is reliable but too heavy for normal iteration.
Category
enhancement: feature
Checklist