> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI agents

> Three agents: the environment assistant builds the world, the control agent drives the robots, the workflow assistant automates work. You approve every plan.

Cyberwave has three agents. They sit on top of everything else: they build the environment, decide what the robots do, and write the automation, all from a plain-language request. Each one shows you what it intends to do and waits for your approval.

<CardGroup cols={3}>
  <Card title="Environment assistant" icon="cubes" href="/ai/environment-assistant">
    **Builds the world.** Describe a scene; it places robots, sensors and objects, and generates assets the catalog doesn't have.
  </Card>

  <Card title="Control agent" icon="gamepad" href="/ai/control-agent">
    **Drives the robots.** Describe a task; it plans actions and skills the twin actually supports and runs the one you approve.
  </Card>

  <Card title="Workflow assistant" icon="diagram-project" href="/ai/workflow-assistant">
    **Automates the work.** Describe a recurring job; it drafts or edits the workflow that does it.
  </Card>
</CardGroup>

## Which one do I need?

| You say                                                      | Agent                 | Result                            |
| ------------------------------------------------------------ | --------------------- | --------------------------------- |
| "A workbench with an SO-101, a camera above it and two bins" | Environment assistant | A scene you can simulate          |
| "Move the Go2 to waypoint A" · "Run the pick skill"          | Control agent         | One approved action, executed now |
| "Inspect every pallet and alert me if one is damaged"        | Workflow assistant    | A workflow that runs on its own   |

A typical session uses all three in order: build the scene, try the task with the control agent in simulation, then turn it into a workflow.

## How approval works

* **Plans come first.** Every agent returns a plan or a proposal before changing anything.
* **You approve one step at a time.** The control agent dispatches a single action only with explicit confirmation. The workflow assistant applies an edit only after **Review and apply**.
* **Simulation before hardware.** Plan and test in simulation, then plan again for the real robot.
* **A person can always take over** from the twin's controller panel. See [Ways to control a robot](/ai/control).

## Where you use them

|                | Environment assistant                    | Control agent                                          | Workflow assistant                               |
| -------------- | ---------------------------------------- | ------------------------------------------------------ | ------------------------------------------------ |
| **Dashboard**  | Environment editor → **Agent**           | Simulate → **Agent**                                   | Workflow workbench → **Prompt** or **Assistant** |
| **Python SDK** | `cw.agents.environment`                  | `cw.agents.control`                                    | `cw.agents.workflow`                             |
| **MCP**        | `cw_plan_scene`, `cw_create_environment` | `cw_plan_control_action`, `cw_dispatch_control_action` | `cw_create_workflow_from_prompt`                 |

To give Claude Code, Cursor or any MCP client these abilities, see the [MCP server](/overview/tools/mcp-server) and the [Claude skill](/overview/tools/claude-skill) under Reference.

## Next steps

<CardGroup cols={2}>
  <Card title="Control agent" icon="gamepad" href="/ai/control-agent">
    The most common starting point: plan, approve and run an action.
  </Card>

  <Card title="Control an SO-101 with your voice" icon="microphone" href="/tutorials/so101-natural-language-agent">
    Voice in, a plan out, the arm moves.
  </Card>
</CardGroup>