AR6420

from GitHub Development & Coding

hydra

Multi-agent orchestration framework for Claude Code. Automatically delegates tasks to cheaper, faster sub-agents (Haiku 4.5, Sonnet 4.6) while maintaining Opus-level quality through verification. Use when working on any coding task — Hydra activates automatically to route file exploration, test running, documentation, code writing, debugging, security scanning, and git operations to the optimal agent. Saves ~50% on API costs. --- # 🐉 Hydra — Multi-Headed Speculative Execution > *"Cut off one head, two more shall take its place."* > Except here — every head is doing your work faster and cheaper. ## Why Hydra Exists Autoregressive LLM inference is memory-bandwidth bound — the time per token scales with model size regardless of task difficulty. Speculative decoding solves this at the token level by having a small "draft" model propose tokens that a large "target" model verifies in parallel. Hydra applies the same principle at the **task level**. Most coding tasks don't need the full reasoning power of Opus. File searches, simple edits, test runs, documentation, boilerplate code — these are "easy tokens" that a faster model handles just as well. By routing them to Haiku or Sonnet heads and reserving Opus for genuinely hard problems, we get: - **2–3× faster task completion** (Haiku responds ~10× faster than Opus) - **~50% reduction in API costs** (Haiku 4.5 is 5× cheaper per token than Opus 4.6) - **Zero quality loss** on tasks within each model's capability band ## How Hydra Works — The Multi-Head Loop ``` User Request │ ├──────────────────────────────────────────────────────┐ │ │ ▼ ▼ ┌─────────────────────────────┐ ┌──────────────────────────────┐ │ 🧠 ORCHESTRATOR (Opus) │ │ 🟢 hydra-scout │ │ Classifies task │ │ IMMEDIATE pre-dispatch: │ │ Plans waves │ │ "Find fil