Using it with agents

• Give the agent a short language guide and the example scripts. They fit in a prompt.
• Let it run sub check in a loop. Errors are worded so a model can fix them.
• Approve with the manifest: compare it to a saved allow-list, or show the plain-words explain output to a person once.
• Centcom can run agent tasks as Subtype scripts and show the permission prompt in its approval flow.

The honest risk

Models write TypeScript and Python better than a new language. The syntax stays close to what they know, and a restricted JSON form of a script can be accepted so tools never have to emit the text syntax.