Give your agent new tools

Tools are your agent's hands. The built-ins already read sensors, drive, and edit files — here you control which it has, and add your own.

Manage the toolbox

1. Panel → bottom TOOLS badge.
2. Checkboxes enable or disable each tool. A disabled tool is invisible to the agent — turning tools off is how you make a safe agent for younger students, or force it to solve a task "blind". The choice is saved in ~/physicar_ws/agent/settings.json, together with the panel's model, effort, context and read-aloud settings — the file appears on your first change (no file means the defaults) and every window on the machine follows it.

Experiment: disable the camera tools and ask "what do you see?" — watch how the agent copes.

Add a custom tool

The fastest way is to ask the tutor itself:

"Add a tool called patrol that drives a slow square and returns to start."

The agent writes the tool into your custom tools file, ~/physicar_ws/agent/tools.py — a plain file in your workspace, next to agent/instructions.md and agent/settings.json, so it survives sessions and travels with a workspace backup. TOOLS → Edit Code opens it yourself. The rules:

• Every top-level function becomes a tool, named custom_<function> for the agent (def patrol → custom_patrol). Names starting with _ are private helpers.
• The docstring is the tool's description; parameters are declared as Annotated[type, Field(description="...")] (typing + pydantic).
• It returns a list of contents — any mix of {"type": "text", "text": ...} and {"type": "image", "mime": ..., "base64": ...}.
• Saving reloads the file; a changed tool is callable from the next message. An import error shows in the TOOLS badge while the last working version keeps serving.

The built-in tools use the exact same format — View Code next to a built-in section shows one as a working example. Once patrol appears in the TOOLS badge (new tools start enabled), you — and the agent — can call it like any built-in.

Good custom tools are small and single-purpose: one action, clear parameters, a short return message. Compose behaviors in instructions, not inside one giant tool.

Try this

Build a report_status tool that returns battery-friendly one-line status (speed + nearest lidar distance), then add an instruction: "After every drive, call report_status."

How tool calling works under the hood → Learn: LLM agents