// SWARMS

Local-first multi-agent orchestration for coding swarms.

Orquestación local-first multiagente para enjambres de código.

Spend intelligence on planning and review. Let deterministic Rust coordinate high-speed, zero-cost, and open-weight agent swarms across isolated Git worktrees.

Invierte inteligencia en planificación y revisión. Deja que un motor determinístico en Rust coordine enjambres rápidos y gratuitos en Git worktrees aislados.

Methodology
SwDD
Swarm-Driven Development. Agents plan, code, and self-verify automatically in parallel branches.
Swarm-Driven Development. Los agentes planean, programan y verifican código automáticamente en paralelo.
Zero-Cost Scaling
Free Tiers
Exploit free quotas of frontier open-weight models to scale rollouts without API cost anxiety.
Aprovecha el uso masivo y gratuito de modelos de frontera. Escala rollouts sin costos de API.
Runtime Engine
100% Rust
Deterministic local orchestration. Isolated Git worktrees & SHA256 integrity locks.
Orquestación local determinística. Git worktrees aislados y verificación SHA256.

Proven scaling: cheap models, frontier accuracy.

Scaling comprobado: modelos económicos, precisión de frontera.

Sampling parallel candidate solutions in detached Git worktrees outperforms expensive proprietary models at up to 11× lower cost.

Muestrear candidatos paralelos en Git worktrees aislados supera a modelos propietarios a un costo hasta 11× menor.

Terminal-Bench 2.1

88% Accuracy · 11× Cheaper 88% Precisión · 11× Más Barato

DeepSeek V4 Flash ×5 with LLM-as-a-Verifier climbs from 79% to 88%, beating Claude Fable 5 while costing ~$0.50/task versus ~$5.50/task.

DeepSeek V4 Flash ×5 con LLM-as-a-Verifier escala de 79% a 88%, superando a Claude Fable 5 costando ~$0.50/tarea frente a ~$5.50/tarea.

~81% Fable 5 (~84%) 79% (Start) DeepSeek ×5 (88%)

DeepSWE Benchmark

87.6% · ~10× Lower Cost 87.6% · ~10× Menor Costo

Empirical scaling from Together AI (Zain Hasan): GLM 5.3 open-weight overtakes Fable at 2-shots (81.1% vs 77.1%) and dominates at 4-shots (87.6% vs 84.1%).

Scaling empírico de Together AI (Zain Hasan): GLM 5.3 supera a Fable en 2-shots (81.1% vs 77.1%) y domina en 4-shots (87.6% vs 84.1%).

2-Shots 77.1% Fable 5 (84.1%) 81.1% GLM 5.3 (87.6%)