Fleet Loop

A multi-agent harness for complex goals. One orchestrator owns the outcome and delegates to specialists. An eval gate blocks slop from compounding. Use when the task is too wide for one agent to do well.

Architecture

Orchestrator
  ├── Research agent     — mines buyer language, market, objections, source
  ├── Strategy agent     — chooses angle, structure, approach
  ├── Specialist agents  — do the narrow work (copy, code, design, analysis)
  ├── Critic agent       — scores output against rubric; rejects weak work
  └── Editor agent       — revises low-scoring sections

An optional Verifier agent runs last to confirm the final output meets the rubric and flags remaining risk.

Prompt template