beastmodelisted

# Beastmode: Multi-Agent Orchestration Framework Beastmode is a structured approach to multi-agent software development that separates high-judgment work (planning, architecture, review) from routine execution (implementation, tests, docs) across different model tiers, with strict cost discipline and a self-improving learning loop. It is designed for MofA (Mixture of Agents) orchestration and pairs well with MemroOS/memroos-style durable memory for long-running agent goals. **Harness-agnostic:** Works with Ultraswarm, GSD, `delegate_task`, Claude Code subagents, or manual orchestration. No specific tool required. ## Core Principle **Frontier models design. Cheap models build and validate. The loop learns.** - **Director/Lead (Design tier):** Frontier model (Claude Fable, Kimi 3, or Opus) owns intent, architecture, creative judgment, and final sign-off. Codex is an explicit opt-in lane only. - **Watcher/Reviewer:** Adversarial reviewer (frontier or mid-tier: Fable/Kimi 3/Opus) challenges plans, gates merges, and catches scope creep. Codex is used here only when explicitly named. - **Executor (Execution tier):** Economy model (MiniMax M3, Qwen 3.7 Plus / Gwen) handles routine implementation *and mechanical validation* (running tests, lint, typecheck, diff summaries) in isolated worktrees. - **Harness:** Any orchestration tool (Ultraswarm, GSD, `delegate_task`, Claude Code subagents, or manual git workflow). - **Memory:** Self-improvement loop records lessons and promotes r