Default light-weight agent workflow with optional handoff to heavy agents and Proctor QA gates.

Standard Flow (Light ↔ Heavy)

The standard Fabrick flow is optimized for interactive work with optional escalation:

1. Overseer builds a TaskShape from the user prompt.
2. Router selects the tier: local (light) or heavy.
3. Worker runs and streams tokens/progress.
4. Proctor validates the output. If it fails, rework or escalate.
5. Auditor records all decisions and events.

Light vs Heavy ownership

• Light (Next.js): fast feedback, low cost, short-lived tasks, great UX.
• Heavy (RunPod/Docker): GPU/vision, long context, batch pipelines, stricter QA.

Decision policy

• Duration forecast and context size → heavy or continuations
• GPU/tools/vision → heavy
• High risk → heavy with stricter Proctor / human-in-loop
• Tight latency → light

flowchart LR
    Prompt --> Overseer --> Router
    Router -->|Local| Light
    Router -->|Heavy| Heavy
    Light --> Proctor
    Heavy --> Proctor
    Proctor -->|Pass| Ready
    Proctor -->|Fail| Rework
    Rework --> Router