How this platform runs

Every run streams a real execution trace — PII scrub, optional tool calls (live web fetch), the model attempt chain, schema validation, and voice composition — as typed spans over SSE. What you see in the theater is the actual pipeline, not an animation.

Outputs are validated against each agent's typed schema before they render; anything that fails validation falls back to a clearly-labeled cached run. Prompts are versioned per agent, and a config eval suite (schema, SEO, Arabic parity, voice-template resolvability, widget wiring) must pass with zero errors before any agent ships.

Open any agent's page and scroll to How it works for its exact system prompt — the library lists all 73.