Skip to main content
Agents plan; people approve. No Cyberwave agent moves a robot, real or simulated, until someone approves that specific action. This page is written for both people and agents. If you are an AI agent, follow the rules below before every action.

​
The rules

  1. Inspect before acting. Read the environment, the twin and its available controls first. With MCP: cw_list_environments, cw_list_twins, cw_list_control_surfaces.
  2. Plan, then wait. Planning never moves anything. Show the plan, including the target twin, direction and limits, and wait for approval.
  3. One bounded action per approval. One pose, one joint target, one navigation goal or one stop. Never chain several motions on one approval.
  4. Simulation first. Run the action in simulation and check the result. Plan again for the real robot.
  5. Verify, don’t assume. An accepted command does not prove motion. Read pose or joint telemetry before and after, and report what was observed.
  6. Don’t retry blindly. If an action times out or its status is unclear, don’t resend it. Read the twin’s state, send a stop if it is safe, and ask the operator.
  7. Ask before destructive changes. Confirm before deleting twins, environments, workflows, models or datasets.

​
What needs approval

​
Built-in guardrails

  • Control dispatch needs confirmation. cw.control.dispatch() defaults to confirmed=False. The MCP tool cw_plan_control_action only plans; cw_dispatch_control_action sends one explicit action.
  • Motion limits. MCP action tools reject navigation over 5.0 m (NAV_DISTANCE_TOO_LARGE) and joint changes over 0.7 rad (JOINT_DELTA_TOO_LARGE). Self-hosted servers can change these limits.
  • Workflows-only environments. If an environment’s control plane is set to workflows only, direct dispatch fails with CONTROL_PLANE_POLICY_DENIED.
  • Preview-first prompts. The MCP server ships safe_navigation, safe_manipulation and inspect_then_act prompts that enforce this flow.

​
Before a live action

Check all four before approving motion on real hardware:
  • The request really is for physical execution, not simulation.
  • Exactly one physical twin is targeted.
  • The twin’s edge device is online and its telemetry is current.
  • A person can stop the robot. Keep the twin’s controller panel open, one click from takeover.
Agents and models can move real hardware. Keep people clear of the robot’s reach while it runs, and test every new task in simulation first.

Control agent

Plan, approve and dispatch from the dashboard or Python.

Teleoperation and takeover

Take control of a twin from the dashboard.