Agent Builder

Agent Builder is the assistant for creating and revising WorldAgents agents.

It follows the same practical pattern as coding agents: inspect the current project state, edit the artifact, use tools, and explain what changed. In WorldAgents, the artifact is an agent plus its supporting files, skills, simulation, and runtime settings.

Agent Builder can help with:

• defining role, goal, voice, greeting, instructions, and boundaries
• attaching files and shared skills
• choosing supported tools and trigger rules
• linking or editing simulation context
• testing behavior in a session and revising the agent

The goal is not to hide the underlying configuration. It is to make the build loop faster while keeping the resulting agent reviewable.

Surfaces

The dashboard builder combines a conversation, full-height Markdown document tabs, and the 3D simulator for staging spaces. Config runs top to bottom: General (name, status, access, session recording), the instruction files, the staged spaces with their recognition setting, and then the collapsed Triggers, Skills, Tools and Agent stack groups. Each heading carries a question mark that explains the group on hover. Triggers and skills are files in the agent's folder; the groups list them and say which need a fix. A System prompt button in the Instructions group opens the prompt the model receives at session start, with platform lines included; it is read-only there and each line links back to the document or setting that produces it. The agent name is a label for people and is not part of the prompt. Config lists the linked instruction files; open each in its own tab to edit the actual text and link supporting files. Markdown stays editable while formatted; plain text and code open in the same lightweight editor from Files. The file menu contains linking and startup-inclusion actions. Files listed as instructions are included at session start; links inside those files are read by the agent when relevant. Runtime settings, tools, and skills remain separate controls. See Agent instructions for file paths and migration.

In chat, type @ or use Browse workspace files to insert a file link at the cursor, such as [style.md](file:/agent/style.md). References stay in the message text and open from the conversation; the builder reads them with its file tools when needed. Uploading a new photo or file still adds it to the message as an upload, and it can be referenced by link in later turns.

The Android app runs the same builder over voice alone. The conversation is the way to make changes, and a read-only card shows the saved name, voice, instruction documents, greeting document, and skills. When the agent is ready to try, the builder hands the conversation over and the app opens a live session with the new draft.

Builder edits the same saved configuration and documents that the dashboard opens. It checks the configuration and linked startup documents after changes and before a test session, then corrects any errors. See Agent file tools for the file and validation workflow.

Agents built this way need no environment. The live camera feed is the environment — whatever the wearer is looking at is what the agent sees — so no space or simulation is staged for them and the builder is told not to ask for one. They also start with the spatial annotation tools enabled, so they can place labels, arrows, and panels into the wearer's view of the room rather than only talking.

What still needs the dashboard: saving a 3D space or simulation, start positions, trigger zones, and publishing. Those tools are left out of the voice builder's tool set, because each of them waits on a 3D view to answer, and the builder is told to say where they live instead of promising them.

Drafts made this way are ordinary DRAFT agents. They appear in the app's agent list and in the dashboard immediately, and publishing one stays a dashboard action.