Agents

A WorldAgents agent is a deployable AI experience for a real workflow.

An agent can include:

• linked instruction documents defining role, goals, tone, and boundaries
• an optional opening instruction document
• a Gemini Live voice configuration
• tools and shared skills
• attached files and workspace context
• trigger rules
• simulation or place context

The API stores agents under the active organization and serves them to the web, iOS, and Android clients. Sessions run through the server-owned realtime runtime.

Status and access

Status describes readiness: Draft, Training, or Live. Access is either organization-only, the default, or public by link.

A Live public agent can be tried without signing in by adults aged 18 and older. The resulting session still belongs to the publishing organization. Public cards include the name and picture, but not instruction document paths or text. Public sessions also disable file tools.

API filtering cannot guarantee that a model will never repeat its instructions. Do not put secrets in an agent prompt.

Templates, samples, and built-ins

• Starter templates are copied into a new agent. They are never run directly.
• Sample agents are public agents that people can run before signing in. They are never copied, and their sessions are never recorded. A sample carries the whole agent, including its instruction documents, triggers, skills, and simulation.
• Built-in agents are product machinery, such as Agent Builder. They are not exposed through public agent routes.

Super admins manage samples in /admin. See Sample agent administration for storage, artwork, access, and session details.

Running an agent

Run opens an agent from the dashboard at /org/<organization id>/agents/<agent id>/run (from the agent card's Run action, or the Run button in Agent Builder). /run/<agent id> runs the same agent with nothing around it: no dashboard navigation, only the running agent, in whichever of your organizations owns it. A run is a live session, the same one the phone and glasses clients open. What the agent looks through is the run's input, picked from the menu on the stage:

• Simulation: a place from the agent's linked simulation, rendered in 3D. Available when the agent has a simulation with environments.
• Camera: the device camera. On a laptop that is the webcam; on a phone it is the phone camera, and the run behaves like the iOS and Android apps.

A laptop opens on the simulation when the agent has one, a phone on the camera, and an agent with no simulation always on the camera. The choice is kept in the link as input=simulation or input=camera, and can be changed during a session. Older /run?agentId=<agent id> and /simulator links redirect to /run/<agent id>.

On a phone the run takes the shape of the iOS and Android session screens: the camera fills the screen, the last few lines of the conversation show over it, and one pill at the bottom holds the input menu, the microphone, the chat (a sheet over the camera) and a menu with session info, the inspector and events. The X in the top bar ends the run and returns to the agents list. On a laptop the dashboard header, the chat rail and the scene controls stay as they are.

This keeps the agent more concrete than a prompt. Teams can edit behavior, update files, run the agent on a simulation or a camera, and reuse the same agent across supported clients.