1. Identity

Fill in the Name — the Agent ID auto-slugs from it until you edit the ID directly. Add a description, set status and visibility, and optionally tag categories. Identity collapses once filled, so it stays out of your way.

When you load an existing agent, the Agent ID is locked — it identifies the record being updated.

2. Modes

A mode is a prompt plus its capabilities. Every agent needs at least one, and one must be the default.

Each mode card collapses to a single line showing its key, prompt handle, and how many things are attached. Expand a mode to edit it — the expanded mode is the one the Library rail adds into.

Set a Prompt by picking a handle, and optionally pin a commit (defaults to latest). Use Preview to read the prompt.

3. Attaching capabilities

The Library rail on the left holds everything available. Drag a chip into the mode, or just click it. Anything already attached leaves the rail, so the rail only ever shows what’s still available.

• Tools — grouped by category, with search and faceted filters (category and usage type) for narrowing a long list.
• Connectors — your connected accounts. Not connected yet? The rail links you to Connections.
• Knowledge — knowledge bases you can access, filterable by tag.
• People — teammates the agent can escalate to.

To remove something, click the × on its chip in the mode.

If clicking a Library chip does nothing, no mode is expanded — expand one first so there’s an unambiguous target.

4. Save

Create agent writes a new agent. If you loaded an existing one, the button becomes Save changes and updates that agent in place, leaving everything you didn’t touch alone.

Show JSON shows the exact record that will be saved.