agent-orchestralisted

# Agent Orchestra A catalog of multi-agent coordination structures and how to map them onto an orchestration runtime. The goal: be *creative and deliberate* about topology — pick (or invent) the structure that fits the task, not reflexively reach for one shape. > **Token-efficient variant:** when budget matters, apply `token-efficient-orchestra` on top of > any graph here — economical models for redundant workers and stronger models for supervision > chokepoints (arbiter/adversary/integrator/gate). The rule of thumb: > *redundancy downgrades, chokepoints upgrade.* --- ## 0. Prime directives 1. **Isolation is sacred.** Independent agents MUST NOT see each other's output. Separate spawns, separate dirs, identical specs. A peek through the wall wastes N−1 agents. 2. **The adversary is always on.** Every produced artifact gets a skeptic whose job is to *refute* it, defaulting to "does not hold" when unsure. This catches failure classes a verifier may pass — a corrupt patch, a vacuous test, a cosmetic timeout. Never ship a pyramid without the refutation stage. 3. **The orchestrator integrates; it does not draft.** You (the caller) apply/merge results. The pyramid's named agents produce content. Never also draft a layer yourself. 4. **Name every agent.** Greek pantheon or any structured convention — never "Worker A". Names make integration legible and the topology readable from a label. 5. **Receipts/provenance.** Every spawn is logged (who, what topology, wh