Pilotfish: Multi-Model Orchestration for Claude Code
Project: Nanako0129/pilotfish — 220⭐, created 2026-07-08 Topics: ai-agents, anthropic, claude, claude-code, multi-agent, orchestration
Summary
A zero-code multi-model orchestration layer for Claude Code: the frontier model (Claude Fable 5 / Opus) plans, decides, and reviews in your main session, while cheaper models execute the volume work. Fresh-context verification guards quality.
Key Metrics
- Anthropic benchmark: Fable 5 orchestrator + Sonnet 5 workers = 96% of all-Fable performance for 46% the cost
- Community audit: Fable orchestrates + Haiku workers = 14.50 (74% savings)
- BrowseComp: 86.8% vs 90.8% accuracy (40.56 per problem)
Architecture
- Three layers: Machine (model alias), Roles (6 agent files), Policy (delegation rules by role, never by model name)
- Six roles: scout (Haiku, read-only), Explore (Haiku), mech-executor (Sonnet), executor (Opus), verifier (Opus, adversarial), security-executor (Opus, off Fable 5)
- Fallback chain:
bestalias auto-resolves;fallbackModel: ["opus", "sonnet"]handles API errors - Key insight: Fresh-context independent verifiers beat self-critique
Relevance to Hermes
Directly parallels our multi-agent architecture (Manas orchestrating Alaya/Anchor/Mano). Verifier pattern could improve cross-modal-review skill. Role-based policy (never naming models) is a cleaner routing pattern.
Bun Rust Rewrite (Bonus Signal)
Source: HN 48837877 | ⬆91pts Blog: https://bun.com/blog/bun-in-rust
Bun (22M+ monthly downloads, Anthropic-owned since Dec 2025) is being rewritten from Zig to Rust. The rewrite addresses extensive memory safety bugs (use-after-free, memory leaks) in the original Zig implementation. The rewrite was performed using pre-release Claude Fable 5.
Impact: Claude Code and OpenCode use Bun as their runtime. Track Rust rewrite milestones for stability improvements.
Andrew Kelley (Zig creator): Acknowledges Zig made Bun possible but criticizes Oven management culture. Highlights the Zig-vs-Rust reliability tradeoff at scale.